A bridge disease intelligent identification method based on an image processing algorithm

Abstract

The application discloses a bridge disease intelligent identification method based on an image processing algorithm and belongs to the technical field of bridge detection and image processing, and comprises the following steps: S1, a consumer-grade unmanned aerial vehicle body is arranged in a closed space in a box girder of a bridge; a multistage protection system is installed outside the unmanned aerial vehicle body; the multistage protection system comprises an elastic buffer shell, a rotor protection cover and a collision buffer support; S2, a light supplement lamp set and a dustproof lens are installed on the unmanned aerial vehicle body; the light supplement lamp set is arranged around the periphery of an image acquisition lens; and the dustproof lens is sealed and covers the front end of the image acquisition lens. Through the cooperative arrangement of the multistage protection system, a heat dissipation protection device, an automatic risk avoidance sensor set, the light supplement lamp set and the dustproof lens, the working safety of the unmanned aerial vehicle in the closed space of the box girder and the image acquisition quality are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of bridge inspection and image processing technology, and more specifically, to a method for intelligent identification of bridge defects based on image processing algorithms. Background Technology

[0002] As a key component of transportation infrastructure, the structural safety of bridges is directly related to public transportation safety and the operation of the national economy. As bridges age, surface defects such as concrete cracks, spalling, exposed reinforcement, water seepage, and honeycomb-like pitting gradually appear. If these defects are not detected and addressed in a timely manner, they may lead to structural performance degradation or even safety accidents.

[0003] The prior art disclosed in CN118506066A is a method for bridge defect detection based on unmanned aerial vehicles (UAVs), comprising the following steps: S100, acquiring historical bridge defect data, the historical bridge defect data including historical bridge defect images and corresponding manual defect annotations, and establishing a historical defect database; S200, using artificial intelligence, training an artificial intelligence model based on the historical bridge defect data in the historical defect database, using historical bridge defect images as input to the input layer and manual defect annotations as output to the output layer, to establish a defect recognition model; S300, acquiring images of the bridge to be detected; S400, using the defect recognition model to analyze the defect data of the bridge images to be detected, generating defect analysis results. This scheme can improve the accuracy and efficiency of bridge defect detection.

[0004] Although the device can improve the accuracy and efficiency of bridge defect detection, traditional bridge defect detection mainly relies on manual visual inspection or bridge inspection vehicle operation. Inspectors need to use telescopes, ladders or lifting platforms to approach the bridge surface for observation and recording. For enclosed and narrow spaces such as the inside of box girders, manual inspection faces difficulties such as limited space, serious dust pollution and poor lighting conditions, resulting in low inspection efficiency and significant safety hazards. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent bridge defect identification method based on image processing algorithms, which solves the aforementioned problems.

[0006] (II) Technical Solution To achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for intelligent identification of bridge defects based on image processing algorithms, comprising the following steps: S1: The consumer-grade drone body is placed in the enclosed space inside the box girder of the bridge. A multi-level protection system is installed on the outside of the drone body. The multi-level protection system includes an elastic buffer shell, a rotor shield and a collision buffer bracket. S2: Install a fill light assembly and a dustproof lens on the drone body. The fill light assembly is arranged around the outer periphery of the image acquisition lens, and the dustproof lens is sealed and covered to the front end of the image acquisition lens. S3: The distance to the inner wall of the box girder is sensed by the automatic avoidance sensor group carried by the UAV body. The automatic avoidance sensor group includes a laser rangefinder, an ultrasonic sensor and an inertial measurement unit. S4: Collect bridge appearance image data, including images of the inside of the box girder and overall images of the bridge exterior; S5: Input the acquired image data into the improved image processing algorithm module, which includes an image preprocessing unit, a feature extraction unit, and a disease classification unit; S6: Output the disease identification results, which include the disease type, location, and geometric dimensions.

[0007] Preferably, the elastic buffer shell is a detachable silicone shell, which covers the outside of the drone body and has a mesh-like reinforcing rib on its surface; the rotor shield is an annular hollow bracket, which is fixed to the periphery of the rotor motor base, and the diameter of the hollow hole in the annular hollow bracket is smaller than the outer diameter of the rotor rotation; the collision buffer bracket includes four arc-shaped elastic legs, which are symmetrically installed on the bottom and top of the drone body, and each arc-shaped elastic leg has a ball bearing at its end.

[0008] Preferably, the number of laser rangefinders is three, with the three laser rangefinders facing the front, rear, and side of the UAV body, respectively; the number of ultrasonic sensors is two, with the two ultrasonic sensors facing the top and bottom of the UAV body, respectively; the inertial measurement unit detects the attitude angle and acceleration of the UAV body in real time; the output terminals of the laser rangefinders, the ultrasonic sensors, and the inertial measurement unit are connected to the flight control module, and the flight control module performs deceleration, hovering, or obstacle avoidance maneuvers based on the sensor output values.

[0009] Preferably, the UAV body is equipped with a heat dissipation protection device, which includes a heat-conducting substrate, two heat pipes, and a micro fan. The heat-conducting substrate is attached to the main control chip and image processing chip of the UAV body. One end of the two heat pipes is connected to the heat-conducting substrate, and the other end of the two heat pipes extends to the side wall of the UAV body. The micro fan is installed inside the side wall of the body, and the air outlet of the micro fan faces the extended ends of the two heat pipes. A heat dissipation grille is provided at a corresponding position on the side wall of the body, and the heat dissipation grille is covered by the elastic buffer shell, which has ventilation holes at corresponding positions.

[0010] Preferably, the image preprocessing unit includes a grayscale conversion subunit, a histogram equalization subunit, a filtering subunit, and a deblurring subunit; the grayscale conversion subunit converts the acquired color image into a grayscale image; the histogram equalization subunit performs adaptive histogram equalization processing on the grayscale image; the filtering subunit removes image noise by using a combination of median filtering and Gaussian filtering; and the deblurring subunit performs Wiener filtering restoration based on point spread function estimation.

[0011] Preferably, the feature extraction unit includes a first scale space subunit, a second feature point detection subunit, a third feature description subunit, and a fourth dimensionality reduction subunit; the first scale space subunit constructs a Gaussian pyramid and a difference-of-Gaussian pyramid for the image; the second feature point detection subunit detects local extrema in the difference-of-Gaussian pyramid and uses the local extrema as candidate feature points; the third feature description subunit calculates the histogram of oriented gradients for each candidate feature point to generate a 128-dimensional feature vector; and the fourth dimensionality reduction subunit uses principal component analysis to reduce the 128-dimensional feature vector to 64 dimensions.

[0012] Preferably, the disease classification unit includes a trained convolutional neural network classifier, the input of which is the feature vector output by the feature extraction unit, and the output of which is the disease category probability distribution; the convolutional neural network classifier includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and a Softmax output layer connected in sequence; the disease categories include cracks, spalling, exposed reinforcement, water seepage, and honeycomb pitting.

[0013] Preferably, it also includes a BeiDou positioning module, which is installed on the top of the UAV body. The BeiDou positioning module collects the spatial position coordinates of the UAV body. The spatial position coordinates are bound and stored with the image acquisition time and image frame number. The spatial position coordinates are output as an attribute field of the defect identification result. When the UAV body is inside the box girder and the BeiDou signal is missing, the relative position coordinates calculated by the inertial measurement unit and the laser rangefinder are used to replace the spatial position coordinates.

[0014] Preferably, the method further includes a bridge external overall condition assessment step, which includes: controlling the UAV to collect all external appearance images of the bridge along a preset flight path; inputting all appearance images into an improved image processing algorithm module to generate a full bridge defect distribution map; calculating the bridge appearance condition index based on the cumulative length of cracks, the cumulative area of ​​spalling areas, the distribution density of exposed rebar points, the cumulative area of ​​seepage areas, and the cumulative area of ​​honeycomb and pitted areas in the defect distribution map; and comparing the bridge appearance condition index with a preset threshold to output the overall bridge assessment level.

[0015] Preferably, the improved image processing algorithm module is integrated into the ground control station; the UAV transmits the collected image data to the ground control station in real time via a wireless communication link; the ground control station runs the image preprocessing unit, the feature extraction unit, and the disease classification unit, and transmits the disease identification results back to the UAV and the handheld terminal; the UAV and the ground control station use frequency hopping spread spectrum communication, with a communication frequency band of 2.4 GHz.

[0016] (III) Beneficial Effects Compared with existing technologies, this invention provides a method for intelligent identification of bridge defects based on image processing algorithms, which has the following beneficial effects: 1. This intelligent bridge defect identification method based on image processing algorithms significantly improves the operational safety and image acquisition quality of UAVs within the enclosed space of box girders through the coordinated configuration of a multi-level protection system, heat dissipation protection device, automatic obstacle avoidance sensor group, supplementary lighting group, and dustproof lens. The elastic buffer shell, rotor guard, and collision buffer bracket absorb impact energy, protect the rotor, and achieve rolling buffering, effectively solving the problem of extremely high collision risk. The heat dissipation protection device efficiently dissipates chip heat through a heat-conducting substrate, heat pipes, and micro fans, resolving the structural imbalance between heat dissipation and collision protection. Laser rangefinders, ultrasonic sensors, and inertial measurement units enable omnidirectional autonomous obstacle avoidance. The supplementary lighting group provides uniform illumination, and the dustproof lens prevents dust from entering, fundamentally improving imaging quality in dimly lit and dusty environments.

[0017] 2. This intelligent bridge defect identification method based on image processing algorithms organically combines an improved image processing algorithm module, a BeiDou positioning module, a bridge external overall condition assessment step, and a ground control station communication architecture to construct a complete UAV bridge inspection application system. The image preprocessing unit removes noise and blur, the feature extraction unit generates a scale-invariant 64-dimensional feature vector, and the convolutional neural network classifier outputs the probability of five types of defects, including cracks and spalling, significantly improving the identification accuracy. The BeiDou positioning module and the inertial measurement unit jointly realize the spatial positioning of defects under all working conditions. Based on five indicators such as crack length and spalling area, the bridge appearance condition index is calculated by weighting and outputting an A to D level rating, providing quantitative technical support for bridge management. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the device of the present invention; Figure 2 This is a diagram illustrating the main steps of the present invention. Detailed Implementation

[0019] The technical solutions of 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 within the scope of protection of the present invention.

[0020] Please see Figures 1-2 The present invention provides a technical solution: A method for intelligent bridge defect identification based on image processing algorithms involves pre-installing a multi-level protection system on the exterior of a consumer-grade drone before it enters the enclosed space inside a bridge box girder. This system includes an elastic buffer shell, rotor guards, and collision buffer supports. The elastic buffer shell is made of detachable silicone material and completely covers the drone's fuselage. When the drone flies within the narrow box girder and collides with its inner wall, the elastic deformation of the silicone absorbs the impact energy, reducing the impact force transmitted to the drone's fuselage. The mesh-like reinforcing ribs on the shell surface increase the shell's structural strength and prevent localized tearing.

[0021] The rotor guard is a ring-shaped perforated bracket fixed to the periphery of the rotor motor mount. The diameter of the perforation in this bracket is smaller than the outer diameter of the rotor's rotation, preventing the rotor from passing through the perforation during high-speed rotation and thus avoiding direct contact between the rotor and the inner wall of the box girder or internal protrusions. When the drone impacts the inner wall of the box girder laterally, the ring-shaped perforated bracket contacts the wall first, distributing the impact force throughout the bracket structure and protecting the rotor blades from damage.

[0022] The collision buffer support consists of four arc-shaped elastic legs, symmetrically mounted on the bottom and top of the UAV body, totaling eight legs. Each arc-shaped elastic leg has a ball bearing at its end. When the UAV impacts the top or bottom plate of the box girder vertically, the arc-shaped elastic legs first contact the impact surface, bending and deforming to absorb the impact energy. The ball bearings then form a rolling contact with the impact surface, converting sliding friction into rolling friction, reducing the risk of jamming between the legs and the wall, and allowing the UAV to slide along the wall and then recover its attitude.

[0023] The drone is equipped with supplementary lighting units and dustproof lenses. The supplementary lighting units are arranged around the image acquisition lens, forming a ring light source. In the dim and dusty environment inside the box girder, the supplementary lighting units emit uniform white light, illuminating the inner wall surface of the box girder to be inspected. The ring arrangement allows the illumination light to be projected from all directions around the lens, reducing shadow areas and specular reflections, and improving the uniformity of illumination on the inspected surface.

[0024] A dustproof lens seals over the front of the image acquisition lens, creating a sealed space between it and the lens housing. Dust particles suspended inside the box girder are dispersed by the airflow from the drone's rotor. When these dust particles come into contact with the dustproof lens, the smooth surface and the sealed structure prevent them from entering the lens's internal optical components. The dustproof lens is made of transparent optical glass with an anti-reflective coating, allowing visible light to pass through efficiently while preventing dust adhesion. After operation, the surface of the dustproof lens can be cleaned to ensure the clarity of subsequently acquired images.

[0025] The drone is equipped with an automatic avoidance sensor suite, including three laser rangefinders, two ultrasonic sensors, and an inertial measurement unit (IMU). The three laser rangefinders are positioned in front of, behind, and to the side of the drone, respectively, to measure the horizontal distance between the drone and the inner wall of the box girder in real time. The two ultrasonic sensors are positioned above and below the drone, respectively, to measure the vertical distance. The IMU monitors the drone's pitch, roll, yaw, and triaxial acceleration in real time.

[0026] The outputs of the laser rangefinder, ultrasonic sensor, and inertial measurement unit (IMU) are connected to the flight control module. The flight control module continuously receives measurement values ​​from each sensor. When the distance measured by the laser rangefinder or ultrasonic sensor in a certain direction is lower than a preset safety threshold, the flight control module determines that there is a collision risk in that direction. If the risk comes from ahead, the flight control module performs a deceleration maneuver to reduce the drone's forward speed. If the risk comes from multiple directions, the flight control module performs a hovering maneuver to maintain the drone's current position. If the distance ahead continues to decrease and deceleration cannot avoid a collision, the flight control module calculates the obstacle avoidance direction based on the attitude angle data provided by the IMU and performs an obstacle avoidance maneuver, controlling the drone to move in the direction with the greater distance.

[0027] The inertial measurement unit (IMU) is also used to compensate for the impact of UAV attitude changes on sensor ranging values. When the UAV tilts, the flight control module performs trigonometric function corrections on the original measurements from the laser rangefinder and ultrasonic sensor based on the pitch and roll angles to obtain the true vertical or horizontal distance.

[0028] The drone's internal structure incorporates a heat dissipation protection device, including a thermally conductive substrate, two heat pipes, and a miniature fan. The thermally conductive substrate is bonded to the drone's main control chip and image processing chip, conducting the Joule heat generated during chip operation to the substrate. One end of each heat pipe is connected to the thermally conductive substrate, and heat from the substrate is transferred to the other end of the heat pipe via a phase-change working fluid inside the heat pipe. The other end of the heat pipe extends to the side wall of the drone's fuselage.

[0029] A miniature fan is installed inside the side wall of the chassis, with its air outlet facing the extensions of the two heat pipes. When powered on, the miniature fan rotates continuously, drawing in cool air from outside the chassis, which flows over the surface of the heat pipe extensions, carrying away heat. The heated air is then exhausted through a heat dissipation grille located at a corresponding position on the side wall of the chassis. The heat dissipation grille is covered by a flexible, cushioned outer shell, which has ventilation holes at corresponding locations to ensure unobstructed airflow while maintaining the shell's cushioning and protective function for the chassis.

[0030] The heat dissipation protection device is activated before the UAV enters the box girder and runs continuously throughout the flight inspection process to ensure that the junction temperature of the main control chip and image processing chip is controlled within the allowable range, preventing the chips from throttling or failing due to overheating.

[0031] During its flight, the drone continuously captures images of the bridge's appearance using its image acquisition lens. The acquired image data includes two categories: images of the box girder's interior and images of the bridge's exterior. Images of the box girder's interior are captured when the drone is inside the enclosed space of the box girder, focusing on the concrete surface of the inner wall. Images of the bridge's exterior are captured when the drone flies over the outside of the bridge, capturing all exposed surfaces, including piers, towers, the sides and bottom of the main beams, and supports.

[0032] The image acquisition lens and the supplementary lighting unit work in tandem. In the dimly lit environment inside the box girder, the image acquisition trigger signal simultaneously activates the supplementary lighting unit, which emits light continuously during exposure and turns off after exposure. In well-lit environments outside the bridge, the supplementary lighting unit does not activate, and imaging relies solely on natural light.

[0033] The aperture, exposure time, and ISO of the image acquisition lens are automatically adjusted according to the ambient light intensity. When the light transmittance of the dustproof lens decreases due to dust accumulation, the automatic exposure algorithm increases the exposure time or increases the ISO to compensate for the loss of light flux.

[0034] The acquired image data is input into an improved image processing algorithm module. This module includes an image preprocessing unit, a feature extraction unit, and a disease classification unit, which work in series.

[0035] The grayscale conversion subunit first converts the acquired color image into a grayscale image. The conversion formula is Gray = 0.299R + 0.587G + 0.114B, where R, G, and B are the red, green, and blue channel values ​​of the color image, respectively. This formula is based on the weighted allocation of human eye sensitivity to different colors, ensuring that the brightness distribution of the converted grayscale image conforms to visual characteristics.

[0036] The histogram equalization subunit performs adaptive histogram equalization on the grayscale image. This process divides the image into multiple 8×8 pixel sub-blocks, calculates the grayscale histogram for each sub-block, and then performs histogram equalization. The boundaries between sub-blocks are smoothly transitioned using bilinear interpolation to avoid block artifacts. This adaptive approach enhances local contrast and addresses the issues of overall image darkness and unclear local details in dimly lit environments inside box girders.

[0037] The filtering subunit employs a cascaded median filtering and Gaussian filtering approach to remove image noise. Median filtering is first applied to the image using a 3×3 pixel sliding window. The median of the sorted grayscale values ​​of all pixels within the window is used as the output value. Median filtering effectively removes salt-and-pepper noise and random impulse noise while preventing excessive smoothing of edge information. The median-filtered image then undergoes Gaussian filtering with a 5×5 pixel kernel and a standard deviation of 1.0 to smooth and suppress any remaining Gaussian noise.

[0038] The deblurring sub-unit performs Wiener filtering restoration based on point spread function estimation. First, the point spread function of the image is estimated using the edge-splitting method, which involves identifying regions with clear edge transitions in the image, analyzing the width and shape of the edge spread, and inferring the point spread function causing the blur. Then, based on the estimated point spread function, Wiener filtering is performed in the frequency domain with a signal-to-noise ratio parameter set to 0.01 to recover the clear image before blurring. This step compensates for motion blur caused by vibration or low-speed movement during UAV flight.

[0039] The first-scale spatial sub-unit constructs the Gaussian pyramid and the difference-of-Gaussian pyramid of the image. The Gaussian pyramid has 4 layers, with 3 scales in each layer, an initial scale of 1.6, and a scale factor of √2. For each layer of the input image, convolution is performed sequentially using Gaussian kernels of different scales to generate Gaussian blurred images at multiple scales within that layer. The Gaussian blurred images at adjacent scales are subtracted to obtain the difference-of-Gaussian pyramid.

[0040] The second feature point detection subunit detects local extrema in the difference-of-gaussian pyramid. For each pixel in the difference-of-gaussian pyramid, it is compared with 26 pixels: 8 pixels in its 3×3 neighborhood at the same scale, 9 pixels at the corresponding position in the previous scale, and 9 pixels at the corresponding position in the next scale. If the gray value of the current pixel is the maximum or minimum value among these 26 neighboring pixels, then the pixel is marked as a candidate feature point.

[0041] The third feature description subunit calculates the oriented gradient histogram for each candidate feature point, generating a 128-dimensional feature vector. First, a 16×16 pixel region is defined centered on the candidate feature point, and this region is divided into 4×4 sub-blocks. Within each sub-block, the gradient magnitude and gradient direction of each pixel are calculated. The 360-degree direction is divided into 8 intervals, and the gradient magnitudes of pixels whose gradient direction falls within that interval are accumulated. Each of the 4×4 sub-blocks generates an 8-dimensional histogram, resulting in a total of 4×4×8=128-dimensional feature vectors.

[0042] The fourth dimensionality reduction subunit uses principal component analysis to reduce the 128-dimensional eigenvectors to 64-dimensionality. A large number of sample eigenvectors are collected, their covariance matrices are calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​are sorted from largest to smallest, and the top 64 eigenvectors with a cumulative contribution rate of 95% are selected to form a projection matrix. The original 128-dimensional eigenvectors are multiplied by this projection matrix to obtain the 64-dimensional reduced eigenvectors. Dimensionality reduction reduces the computational burden of subsequent classifiers while retaining the main discriminative information.

[0043] The disease classification unit includes a trained convolutional neural network classifier. The classifier's input is a 64-dimensional feature vector output from the feature extraction unit, and its output is the probability distribution of disease categories. The network structure of the convolutional neural network classifier consists of, in sequence, an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and a softmax output layer.

[0044] The input layer receives a 64-dimensional feature vector and reshapes it into an 8×8 feature map. The first convolutional layer uses 32 3×3 pixel convolutional kernels with a stride of 1 pixel to convolve the input feature map, outputting 32 feature maps. The first pooling layer uses a 2×2 pixel max pooling window with a stride of 2 pixels to downsample the output of the first convolutional layer, halving the feature map size. The second convolutional layer uses 64 3×3 pixel convolutional kernels with a stride of 1 pixel to convolve the output of the first pooling layer, outputting 64 feature maps. The second pooling layer uses a 2×2 pixel max pooling window with a stride of 2 pixels for further downsampling. The fully connected layer unfolds all neurons output from the second pooling layer into a one-dimensional vector, and performs feature fusion through a 128-neuron fully connected layer. The Softmax output layer maps the 128-dimensional output of the fully connected layer into a 6-dimensional probability vector, corresponding to six categories: cracks, peeling, exposed reinforcement, water seepage, honeycomb texture, and background, with a sum of probabilities of 1. The category with the highest probability is output as the disease identification result for that image patch.

[0045] The BeiDou positioning module is installed on the top of the UAV. When flying outside the bridge, the BeiDou positioning module receives BeiDou satellite signals and calculates the UAV's spatial position coordinates in real time, including longitude, latitude, and altitude. These spatial position coordinates are stored in conjunction with the image acquisition time and image frame number.

[0046] The image acquisition time and image frame number are generated by the main control chip of the UAV. For each image frame acquired, the main control chip reads the current system time as the acquisition time and increments the frame number counter by 1. The BeiDou positioning module outputs position coordinates using the same time base. The main control chip associates the position coordinates with the image frame closest to the acquisition time, forming a complete data record containing image data, position coordinates, timestamp, and frame number.

[0047] When the drone enters the box girder, the BeiDou signal is blocked by the concrete structure, causing positioning failure. At this point, the relative position coordinates calculated jointly by the inertial measurement unit (IMU) and the laser rangefinder are used instead of the BeiDou coordinates. Using the last valid BeiDou coordinates at the box girder entrance as the origin, the IMU performs a double integration of the acceleration data to obtain the displacement change from the origin, thus calculating the drone's relative position at the current moment. The laser rangefinder measures the distance between the drone and the inner wall of the box girder; this distance constraint is compared with the known geometric dimensions of the box girder, and the accumulated error from the inertial integration is periodically corrected.

[0048] The relative position coordinates are output as an attribute field of the defect identification result, so that each identified defect can be located in a specific spatial position inside the box girder.

[0049] After acquiring all external images of the bridge, the overall external condition assessment is performed. First, the drone is controlled to automatically fly along a preset flight path to collect images. The preset flight path is an S-shaped path with a spacing of 2 to 3 meters, ensuring a 60% to 80% overlap between images of adjacent paths. The flight speed is controlled at 0.5 to 1 meter per second to ensure image clarity.

[0050] All acquired surface images are sequentially input into the improved image processing algorithm module for frame-by-frame defect identification. After identification, using the bridge's 3D model as a base, the identification result for each defect is labeled at its corresponding location within the model. The labeling includes defect type (cracks, spalling, exposed reinforcement, water seepage, honeycomb surface defects), defect size (crack length, spalling area, etc.), and defect confidence level (probability value output by the convolutional neural network). All labeled results are then aggregated to generate a full-bridge defect distribution map.

[0051] The bridge appearance condition index is calculated based on the overall bridge defect distribution map. The calculation formula is: BCI = 100 × (1 - w1×Lc / Ltotal - w2×Ac / Atotal - w3×ρr - w4×Aw / Atotal - w5×Ah / Atotal) Wherein, Lc is the cumulative length of cracks across the entire bridge (in meters), Ltotal is the characteristic length corresponding to the total surface area of ​​the bridge (in meters), Ac is the cumulative area of ​​the spalling region (in square meters), Atotal is the total surface area of ​​the bridge (in square meters), ρr is the distribution density of exposed rebar points (in units / square meter), Aw is the cumulative area of ​​the seepage region (in square meters), and Ah is the cumulative area of ​​the honeycomb and pitted surface region (in square meters). The weighting coefficients w1 to w5 are 0.3, 0.25, 0.2, 0.15, and 0.1, respectively.

[0052] After calculating the Bridge Appearance Quality Index (BCI), it is compared with preset thresholds. The preset thresholds include three cutoff values: 85, 70, and 55. When the BCI is greater than or equal to 85, the overall bridge rating is output as Grade A (Good Condition). When the BCI is greater than or equal to 70 but less than 85, Grade B (Slight Damage) is output. When the BCI is greater than or equal to 55 but less than 70, Grade C (Moderate Damage) is output. When the BCI is less than 55, Grade D (Severe Damage) is output.

[0053] The improved image processing algorithm module is integrated into the ground control station and does not run on the UAV itself, thus reducing the UAV's computational load and power consumption. The UAV transmits the acquired image data to the ground control station in real time via a wireless communication link. The wireless communication link operates at a frequency of 2.4 GHz and has a transmission rate of 10 Mbps to 50 Mbps, which meets the bandwidth requirements for real-time transmission of high-definition images.

[0054] The UAV and the ground control station communicate using frequency-hopping spread spectrum. The frequency-hopping sequence is a pseudo-random sequence, with a hopping rate of 1000 hops per second and each hop lasting 1 millisecond. Both communicating parties synchronously change carrier frequencies according to the same pseudo-random sequence, evenly distributed within the 2.4 GHz band. Frequency-hopping spread spectrum improves anti-jamming and anti-interception capabilities, avoiding co-channel interference when multiple UAVs operate simultaneously.

[0055] After receiving the image data, the ground control station sequentially runs the image preprocessing unit, feature extraction unit, and disease classification unit to complete disease identification. The identification results include the type, location coordinates, and geometric dimensions of the disease detected in each image frame. The ground control station transmits the disease identification results back to the UAV via the same wireless communication link, and simultaneously sends them to the handheld terminal via Bluetooth or WiFi.

[0056] The handheld terminal displays the damage identification results and the drone's flight status in real time. Flight status includes the drone's current position, attitude angle, battery level, and sensor readings. Operators can monitor the detection progress, manually mark suspected damage areas, or adjust the drone's flight path using the handheld terminal.

[0057] After receiving the disease identification results, the UAV associates and stores the results with the original image data stored on its own to form a complete detection record. If the wireless communication link is temporarily interrupted, the UAV caches the image data and flight data in its local storage and automatically resumes transmission when communication is restored.

[0058] Operators placed the drone, equipped with a multi-level protection system, supplementary lighting, dustproof lenses, automatic obstacle avoidance sensors, heat dissipation protection devices, and a BeiDou positioning module, at the entrance of the bridge box girder. The drone was started, and the flight control module performed sensor self-checks and initial positioning. The drone entered the box girder, and the flight control module automatically avoided obstacles based on real-time ranging values ​​from laser and ultrasonic sensors. With the assistance of supplementary lighting, the image acquisition lens continuously captured images of the box girder's inner wall, and the image data was transmitted in real-time to the ground control station via a frequency-hopping spread spectrum communication link. The ground control station sequentially performed image preprocessing, feature extraction, and defect classification, outputting the defect identification results and transmitting them back. After completing the internal inspection of the box girder, the drone flew out of the box girder, switched to BeiDou positioning mode, and collected images of the entire external appearance of the bridge along a preset S-shaped flight path, transmitting and processing them in real-time as well. After all inspections were completed, the ground control station generated a full-bridge defect distribution map, calculated the bridge appearance condition index, and output the overall assessment level. Operators viewed the inspection report through a handheld terminal, located the defect locations, and assessed the current state of the bridge structure.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of bridge defects based on image processing algorithms, characterized in that, Includes the following steps: S1: The consumer-grade drone body is placed in the enclosed space inside the box girder of the bridge. A multi-level protection system is installed on the outside of the drone body. The multi-level protection system includes an elastic buffer shell, a rotor shield and a collision buffer bracket. S2: Install a fill light assembly and a dustproof lens on the drone body. The fill light assembly is arranged around the outer periphery of the image acquisition lens, and the dustproof lens is sealed and covered to the front end of the image acquisition lens. S3: The distance to the inner wall of the box girder is sensed by the automatic avoidance sensor group carried by the UAV body. The automatic avoidance sensor group includes a laser rangefinder, an ultrasonic sensor and an inertial measurement unit. S4: Collect bridge appearance image data, including images of the inside of the box girder and overall images of the bridge exterior; S5: Input the acquired image data into the improved image processing algorithm module, which includes an image preprocessing unit, a feature extraction unit, and a disease classification unit; S6: Output the disease identification results, which include the disease type, location, and geometric dimensions.

2. The method according to claim 1, characterized in that, The elastic buffer shell is a detachable silicone shell that covers the outside of the drone body. The surface of the silicone shell is provided with a grid-like reinforcing rib. The rotor guard is an annular hollow bracket that is fixed to the periphery of the rotor motor base. The diameter of the hollow hole in the annular hollow bracket is smaller than the outer diameter of the rotor rotation. The collision buffer bracket includes four arc-shaped elastic legs that are symmetrically installed on the bottom and top of the drone body. Each arc-shaped elastic leg has a ball bearing at its end.

3. The method according to claim 1, characterized in that, The system comprises three laser rangefinders, which are positioned facing the front, rear, and side of the UAV body, respectively. It also comprises two ultrasonic sensors, which are positioned facing the top and bottom of the UAV body, respectively. The inertial measurement unit (IMU) monitors the UAV body's attitude angle and acceleration in real time. The outputs of the laser rangefinders, ultrasonic sensors, and IMU are connected to the flight control module, which performs deceleration, hovering, or obstacle avoidance maneuvers based on the sensor output values.

4. The method according to claim 1, characterized in that, The drone body is equipped with a heat dissipation protection device, which includes a heat-conducting substrate, two heat pipes, and a miniature fan. The heat-conducting substrate is attached to the main control chip and image processing chip of the drone body. One end of the two heat pipes is connected to the heat-conducting substrate, and the other end of the two heat pipes extends to the side wall of the drone body. The miniature fan is installed inside the side wall of the body, and the air outlet of the miniature fan faces the extended ends of the two heat pipes. A heat dissipation grille is provided at a corresponding position on the side wall of the body. The heat dissipation grille is covered by the elastic buffer shell, and the elastic buffer shell has ventilation holes at corresponding positions.

5. The method according to claim 1, characterized in that, The image preprocessing unit includes a grayscale conversion subunit, a histogram equalization subunit, a filtering subunit, and a deblurring subunit; the grayscale conversion subunit converts the acquired color image into a grayscale image; The histogram equalization subunit performs adaptive histogram equalization processing on the grayscale image; The filtering subunit removes image noise by using a combination of median filtering and Gaussian filtering; the deblurring subunit performs Wiener filtering restoration based on point spread function estimation.

6. The method according to claim 1, characterized in that, The feature extraction unit includes a first scale space subunit, a second feature point detection subunit, a third feature description subunit, and a fourth dimensionality reduction subunit. The first scale space subunit constructs a Gaussian pyramid and a difference-of-Gaussian pyramid for the image. The second feature point detection subunit detects local extrema in the difference-of-Gaussian pyramid and uses these local extrema as candidate feature points. The third feature description subunit calculates the histogram of directional gradients for each candidate feature point, generating a 128-dimensional feature vector. The fourth dimensionality reduction subunit uses principal component analysis to reduce the 128-dimensional feature vector to 64 dimensions.

7. The method according to claim 1, characterized in that, The disease classification unit includes a trained convolutional neural network classifier. The input of the convolutional neural network classifier is the feature vector output by the feature extraction unit, and the output of the convolutional neural network classifier is the probability distribution of disease categories. The convolutional neural network classifier includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and a Softmax output layer connected in sequence. The disease categories include cracks, spalling, exposed reinforcement, water seepage, and honeycomb pitting.

8. The method according to claim 1, characterized in that, It also includes a BeiDou positioning module, which is installed on the top of the UAV body. The BeiDou positioning module collects the spatial position coordinates of the UAV body. The spatial position coordinates are bound and stored with the image acquisition time and image frame number. The spatial position coordinates are output as an attribute field of the defect identification result. When the UAV body is inside the box girder and the BeiDou signal is missing, the relative position coordinates calculated by the inertial measurement unit and the laser rangefinder are used to replace the spatial position coordinates.

9. The method according to claim 1, characterized in that, It also includes a bridge external overall condition assessment step, which includes: controlling the UAV to collect all external appearance images of the bridge along a preset flight path; inputting all appearance images into an improved image processing algorithm module to generate a full bridge defect distribution map; calculating the bridge appearance condition index based on the cumulative length of cracks, the cumulative area of ​​spalling areas, the distribution density of exposed rebar points, the cumulative area of ​​seepage areas, and the cumulative area of ​​honeycomb and pitted areas in the defect distribution map; comparing the bridge appearance condition index with a preset threshold and outputting the overall bridge assessment level.

10. The method according to claim 1, characterized in that, The improved image processing algorithm module is integrated into the ground control station; the UAV transmits the collected image data to the ground control station in real time via a wireless communication link; the ground control station runs the image preprocessing unit, the feature extraction unit, and the disease classification unit, and transmits the disease identification results back to the UAV and the handheld terminal; the UAV and the ground control station use frequency hopping spread spectrum communication, with a communication frequency band of 2.4 GHz.

Images