Steel surface damage identification device and identification method based on binocular machine vision of visible light image and infrared image

Abstract

The invention provides a steel surface damage identification device based on binocular machine vision of a visible light image and an infrared image, which is characterized by comprising a visible light image shooting device, an infrared image shooting device, a guide rail, a controller and data storage equipment. The invention also provides a steel surface damage identification method based on binocular machine vision of the visible light image and the infrared image, which comprises the following steps: firstly, preliminarily identifying apparent damage by using a traditional camera, and determining a suspected area; then, an infrared camera is adopted to collect an infrared thermal image, and whether the area is an apparent damage area or a suspected area is determined again; and finally, acquiring an image of the suspected area after the high-pressure water mist is sprayed by adopting a traditional camera, and performing final damage type determination by combining the images of the infrared camera and the traditional camera and utilizing a pre-trained neural network model, so that the defect type is identified more accurately and more efficiently.

Description

Technical Field

[0001] The present invention relates to the technical field of steel surface damage, and in particular to a steel surface damage identification device and an identification method. Background Art

[0002] Surface damage to steel structures significantly impacts the overall performance and safe operation of cranes. Specific hazards are as follows: Surface damage, such as rust, can lead to uniform thinning of the steel structure and the creation of large localized rust pits. This not only easily causes stress concentration but can also lead to premature localized structural failure. Rust can significantly reduce the fatigue allowable stress and cold brittleness resistance of steel structures subjected to direct dynamic loads and those exposed to low temperatures. Excessive rust not only reduces crane performance but can also cause crane collapse during operation, potentially resulting in serious casualties. Furthermore, loosening of the tower section connecting bolts is a common form of damage to tower crane steel structures and a major potential safety hazard. Loose bolts directly impact the stability and overall strength of the tower structure and, in severe cases, can lead to overall structural failure. Therefore, research on surface damage in tower crane steel structures is of paramount importance. By thoroughly studying the mechanisms and identification methods of surface damage, crane accidents can be effectively predicted and prevented, ensuring the safety of personnel and property. This can also extend the crane's service life and operating efficiency, while reducing operational and maintenance costs.

[0003] By establishing an accurate and reliable surface damage identification method, surface damage on steel structures can be detected promptly and accurately, allowing appropriate repair and reinforcement measures to be implemented. This not only prevents further damage but also avoids safety accidents caused by failure to detect and address damage in a timely manner. In recent years, machine vision-based surface damage detection methods have been widely used for surface damage detection of steel structures. These methods offer the following significant advantages: 1) high precision and efficiency. Using image processing methods, they can accurately identify subtle surface damage on steel structures, such as cracks, rust, and deformation. Furthermore, machine vision-based methods can rapidly process large amounts of image data, enabling fast and efficient surface damage detection. 2) non-contact and automated. Traditional damage detection methods require close observation and manual measurement, which are not only time-consuming and labor-intensive, but also pose certain safety risks. Machine vision-based detection methods avoid direct contact with the structure being inspected. Furthermore, when combined with technologies such as drones, they can automate the collection of surface damage images, effectively avoiding errors and damage caused by human factors while reducing the tediousness and labor intensity of manual operations. 3) high stability and reliability. Machine vision-based inspection methods typically possess the ability to self-learn and optimize, improving detection accuracy and stability through continuous learning and data accumulation. 4) Implementing data-based and visual inspection. Machine vision methods capture surface images of structures and can display and record detected damage information in the form of images and data, facilitating subsequent data analysis and processing.

[0004] However, the detection method based on machine vision also has the following disadvantages: 1) The machine vision system relies heavily on the quality of the image, which may be affected by environmental factors such as lighting, temperature, and humidity. For example, insufficient lighting or uneven lighting may cause blurred images or unclear features, thus affecting the accuracy of damage detection. 2) The surface of steel is usually coated, and different coatings have different reflection and scattering properties, resulting in different performance during detection. In addition, some structural surfaces may have low contrast or complex textures, which makes damage features difficult to extract and identify. 3) For some complex forms of damage, such as deep damage, coexistence of multiple types of damage, etc., traditional machine vision detection methods may be difficult to accurately identify and classify. Summary of the Invention

[0005] To address the above technical issues, the present invention proposes a device and method for identifying steel surface damage based on binocular machine vision using visible light and infrared images. This method uses an infrared camera to capture images based on the temperature difference between surface defects and intact parts of the steel. Combined with images captured by traditional cameras under visible light, deep learning methods are used to establish a correspondence between traditional and infrared images and the type of apparent damage. The method proposed in this invention can more accurately and efficiently identify defect types, thereby effectively monitoring and assessing the health status and potential problems of the structure, providing a more comprehensive and accurate service for evaluating the service performance of steel structures. Furthermore, this method can provide an important basis for formulating reasonable maintenance and overhaul plans, thereby reducing the maintenance costs and risks of steel structures.

[0006] In order to achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] A steel surface damage identification device based on binocular machine vision of visible light images and infrared images includes a visible light image shooting device, an infrared image shooting device, a guide rail, a controller and a data storage device. The visible light image shooting device and the infrared image shooting device can be slidably arranged on one side of the guide rail, and the controller and the data storage device are fixed on the other side of the guide rail. The visible light image shooting device and the infrared image shooting device are both wirelessly connected to the controller, and the data storage device is connected to the controller.

[0008] Specifically, the visible light image capturing device includes a first sliding base, a conventional camera and a white light source. The conventional camera and the white light source are both located on a side of the first sliding base away from the guide rail, and the conventional camera is located on a side of the white light source.

[0009] Specifically, the infrared image shooting device includes a second sliding base, an infrared camera and an infrared light source. The infrared camera and the infrared light source are both located on a side of the second sliding base away from the guide rail, and the infrared light source is arranged on one side of the infrared camera.

[0010] Specifically, a positioning module I and a positioning module II are respectively provided on a side of the first sliding base close to the guide rail and a side of the second sliding base close to the guide rail.

[0011] Specifically, the first sliding base is provided with a wireless receiving module I, a motor driving circuit I and a stepping motor I connected in sequence; the second sliding base is provided with a wireless receiving module II, a motor driving circuit II and a stepping motor II connected in sequence, and the wireless receiving module I and the wireless receiving module II are both connected to the controller.

[0012] Specifically, the second sliding base is further provided with a high-pressure spray system, which is arranged on the other side of the infrared camera.

[0013] A method for identifying steel surface damage based on binocular machine vision of visible light images and infrared images comprises the following steps:

[0014] S1: calibrate the surface feature data of the undamaged steel to obtain normal statistical values ​​of visible light image pixel values ​​and normal statistical values ​​of infrared image pixel values ​​of the undamaged steel;

[0015] S2: Determine the apparent damaged area and the suspected damaged area using a visible light image capturing device and an infrared image capturing device according to the normal statistical values ​​of the visible light image pixel values ​​and the normal statistical values ​​of the infrared image pixel values, and obtain a visible light image of the suspected damaged area and an infrared image of the suspected damaged area;

[0016] S3: Preprocessing the visible light image and the infrared image of the suspected damaged area, performing image registration and cropping, and performing statistical analysis to obtain a probability density distribution map of the visible light image and a probability density distribution map of the infrared image;

[0017] S4: Using the registered and cropped visible light image of the suspected damage area, the probability density distribution map of the visible light image, and the infrared image of the suspected damage area, the probability density distribution map of the infrared image as input, respectively, a pre-trained generative adversarial network is used for classification to determine whether the suspected damage area is a damage area and the damage type.

[0018] Furthermore, the implementation method of step S2 is:

[0019] S2.1: The controller controls the traditional camera to move to a position above the steel surface, turns on a white light source, illuminates area A on the steel surface, and uses the traditional camera to capture a visible light image of the steel. The controller detects whether any pixel values ​​in the visible light image deviate from normal statistical values ​​of visible light image pixel values. If so, the controller identifies area A on the steel surface as a suspected damage area, obtains a visible light image of the suspected damage area, and proceeds to step S2.2. Otherwise, the controller moves to the next area on the steel surface for inspection.

[0020] S2.2: After the controller turns off the white light source, it controls the first sliding base to move the white light source and the traditional camera away, and controls the second sliding base to move the high-pressure spray system, the infrared light source, and the infrared camera to area A on the steel surface;

[0021] S2.3: The controller controls the high-pressure spray system to start, spraying high-pressure water mist to clean area A on the surface of the steel;

[0022] S2.4: The controller controls the infrared light source to turn on, illuminate area A on the steel surface for a few seconds, and then uses the infrared camera to capture an infrared image of area A on the steel surface;

[0023] S2.5: Crop the visible light image and infrared image of area A on the steel surface separately, retain the image of the intersection area, and compare the infrared image of the intersection area. If the deviation from the normal statistical value of the infrared image pixel value exceeds the threshold, it is determined to be an apparent damage area; otherwise, it is still a suspected damage area.

[0024] Furthermore, the preprocessing method is as follows: denoising, contrast enhancement, and smoothing filtering are performed on the collected visible light image of the suspected damage area; temperature calibration and pseudo-color mapping are performed on the infrared image of the suspected damage area; and the SIFT algorithm is used to extract extreme points in the visible light image and the infrared image of the suspected damage area as feature points of the image.

[0025] Furthermore, the method for image registration is as follows: a rectangular coordinate system is established for each of the preprocessed conventional image and the infrared image, a square consisting of four pixels is defined as a pixel block, and the coordinates (x0, y0) of each pixel block in the preprocessed conventional image rectangular coordinate system and the coordinates (x, y) of each pixel block in the preprocessed infrared image rectangular coordinate system are given; the relative rotation angle is θ, and the relative position translation vector is (a, b). The position formula of the infrared image becomes:

[0026]

[0027] Change the relative rotation angle to θ and the relative position translation vector (a, b) to maximize (x0·x'+y0·y'). At this time, the corresponding relative rotation angle is θ, and the relative position translation vector (a, b) aligns the feature points in the infrared image and the traditional image in space.

[0028] The beneficial effects of the present invention are:

[0029] 1. Combining the advantages of traditional cameras and infrared cameras, it can simultaneously obtain visible light information and infrared information of the steel surface, improving the accuracy of damage identification;

[0030] 2. Using statistical methods to compare density probability distribution maps to classify damage degree and damage area, providing an objective and quantitative basis for steel quality monitoring;

[0031] 3. Through feature extraction and classification recognition, accurate identification of various types of steel surface damage can be achieved;

[0032] 4. Use high-pressure spray to spray water mist to clean the surface of steel to avoid identification errors caused by surface attachments;

[0033] 5. Combining the high resolution of traditional cameras with the thermal imaging characteristics of infrared cameras, it can accurately capture subtle damage and temperature changes on the steel surface, thereby improving the accuracy of damage identification;

[0034] 6. Infrared cameras can sense the temperature distribution on the surface of steel. For local temperature changes caused by cracks, rust, etc., infrared cameras can quickly capture these abnormal phenomena and effectively identify the damage;

[0035] 7. To ensure the stability of the temperature field on the steel surface, use an infrared light source to heat the steel surface before using the infrared camera to collect infrared thermal images to avoid recognition errors caused by temperature field changes;

[0036] 8. This method is not only suitable for damage identification on the surface of steel, but can also be applied to damage detection of other metal materials, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 Schematic diagram of the structure of the identification device of the present invention.

[0039] Figure 2 Pixel maps of damaged and non-damaged areas.

[0040] Figure 3 This is a flow chart of the identification method of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0042] Example 1

[0043] A steel surface damage recognition device based on binocular machine vision of visible light images and infrared images, such as Figure 1As shown. It includes a visible light image capture device, an infrared image capture device, a guide rail 1, a controller, and a data storage device. The visible light image capture device and the infrared image capture device can both be slidably mounted on one side of the guide rail 1, while the controller and the data storage device are fixed to the other side of the guide rail 1. Both the visible light image capture device and the infrared image capture device are wirelessly connected to the controller, and the data storage device is connected to the controller. Guide rail 1 supports the sliding bases of the visible light image capture device and the infrared image capture device, providing a horizontal motion track to enable alternating scanning and position switching between the two cameras. It utilizes HIWIN EG series linear guides, which offer high load capacity and strong wear resistance. The controller, a Raspberry Pi 5, controls the operation of the visible light image capture device and the infrared image capture device and processes image data.

[0044] The visible light image capture device includes a first sliding base 3, a conventional camera 5, and a white light source 6. Both the conventional camera 5 and the white light source 6 are located on the side of the first sliding base 3 away from the guide rail 1, with the conventional camera 5 located on the side of the white light source 6. The conventional camera 5, a Basler ace 2 series camera, is used to capture RGB images of the steel surface. The white light source 6, a CCS LDR2-100W series ring-shaped LED light source, provides uniform illumination, eliminates ambient light interference, and enhances image contrast.

[0045] The infrared image capture device includes a second sliding base 11, an infrared camera 8, and an infrared light source 10. Both are located on the side of the second sliding base 11 away from the guide rail 1, with the infrared light source 10 positioned to one side of the infrared camera 8. The infrared camera 8, a FLIR A65, is used to detect the surface temperature distribution of the steel and capture local thermal conductivity differences caused by damage. The infrared light source, an Opto Engineering TC series, is used to actively heat the steel surface, stimulating infrared signatures of potentially damaged areas through thermal radiation and preventing recognition errors caused by temperature field variations.

[0046] Positioning Modules I and II are installed on the side of the first sliding base 3 near the guide rail 1 and the side of the second sliding base 11 near the guide rail 1, respectively. The guide rail is marked with a scale. Positioning Modules I and II record the sliding position of the camera on the guide rail. Based on the positioning system and scale, the camera position is determined, and the damage location is then determined. Both modules use Omron E6B2-CWZ6C photoelectric rotary encoders.

[0047] The first sliding base 3 is equipped with a wireless receiving module I, a motor drive circuit I, and a stepper motor I, which are connected in sequence. The second sliding base 11 is equipped with a wireless receiving module II, a motor drive circuit II, and a stepper motor II, which are connected in sequence. Both wireless receiving modules I and II are connected to a controller. Wireless receiving modules I and II are used to receive controller commands and reduce cable interference. Both utilize the ESP32-WROOM-32E. The stepper motor and motor drive circuit control the movement of the sliding base along the guide rails to achieve precise positioning. They utilize a Lesai intelligent DM542 driver and a 57HS22 stepper motor. In this embodiment, the stepper motor directly engages with the gear via a D-shaped shaft, driving the gear to rotate and thereby drive the first sliding base 3 to move.

[0048] The second sliding base 11 is also provided with a high-pressure spray system 9, which is provided on the other side of the infrared camera 8. The high-pressure spray system 9 is used to clean surface stains and eliminate the interference of the oxide layer on thermal imaging. It uses a SprayingSystems 1 / 4J series fan nozzle.

[0049] Example 2

[0050] A steel surface damage identification method based on binocular machine vision of visible light images and infrared images, the steps are as follows:

[0051] S1: calibrate the surface feature data of the undamaged steel to obtain normal statistical values ​​of visible light image pixel values ​​and normal statistical values ​​of infrared image pixel values ​​of the undamaged steel.

[0052] The method for calibrating the surface characteristic data of undamaged steel is as follows:

[0053] S1.1: Illuminate the surface of each type of undamaged steel material using a white light source 6 to obtain a visible light image of the surface of the undamaged steel material. Statistically calculate the distribution of pixel values ​​in the R, G, and B matrices. Fit the pixel values ​​of each R, G, and B matrix using a normal distribution to obtain normal statistical values ​​of the pixel values ​​in the visible light image of each type of undamaged steel material, which are the mean and variance of each pixel matrix.

[0054] S1.2: Use a high-pressure spray system 9 to clean the undamaged steel surface. After cleaning, use an infrared light source 10 to irradiate for 10 seconds. Then, use an infrared camera 8 to capture an infrared thermal image. The distribution of pixel values ​​in the temperature value matrix of the infrared thermal image is statistically analyzed to obtain the normal statistical values ​​(mean and variance) of the infrared image pixel values ​​of the undamaged steel surface.

[0055] S2: Determine the apparent damage area and the suspected damage area using a visible light image capturing device and an infrared image capturing device according to the normal statistical values ​​of the visible light image pixel values ​​and the normal statistical values ​​of the infrared image pixel values, and obtain a visible light image of the suspected damage area and an infrared image of the suspected damage area.

[0056] S2.1: The controller controls the traditional camera 5 to move to a position above the steel surface, turns on the white light source 6, illuminates the steel surface area A, uses the traditional camera 5 to collect the visible light image of the steel, and detects whether there is a pixel value in the visible light image that deviates from the normal statistical value of the visible light image pixel value. If deviated, the steel surface area A is determined as a suspected damage area, and the visible light image of the suspected damage area is obtained, and step S2.2 is performed. Otherwise, move to the next area on the steel surface for detection.

[0057] Method for judging whether the visible light image pixel value deviates from the normal statistical value: compare the measured pixel value in the visible light image with the pixel mean of various types of undamaged steel. If the error with the mean is less than 3 times the variance, it indicates a good match, otherwise it does not match; if it does not match with all materials, it indicates that the pixel point is abnormal.

[0058] Specifically, in this embodiment, if more than 10 connected pixels deviate from the normal value by more than 3 times the variance, and at least 2 pixels are connected (as shown in the figure), it indicates that the area is abnormal. Scattered abnormal pixels are abnormal pixel maps of non-damaged areas, such as Figure 2 shown.

[0059] S2.2: After the controller controls the white light source 6 to be turned off, it controls the first sliding base 3 to move the white light source 6 and the traditional camera 5 away, and controls the second sliding base 11 to move the high-pressure spray system 9, the infrared light source 10 and the infrared camera 8 to the steel surface area A.

[0060] S2.3: The controller controls the high-pressure spray system 9 to start, spraying high-pressure water mist to clean the surface area A of the steel material.

[0061] S2.4: The controller controls the infrared light source 10 to turn on, and after irradiating the steel surface area A for a few seconds, an infrared camera is used to collect an infrared image of the steel surface area A.

[0062] S2.5: Crop the visible light image and infrared image of area A on the steel surface separately, retain the image of the intersection area, and compare the infrared images of the intersection area. If the pixel values ​​deviate from the normal statistical value of the infrared image, it is determined to be an apparent damage area; otherwise, it is still a suspected damage area.

[0063] The method for judging whether the infrared image pixel value deviates from the normal statistical value is as follows: the measured pixel value in the infrared image is compared with the pixel mean of various types of undamaged steel. If the error with the mean is less than 3 times the variance, it indicates a good match; otherwise, it does not match; if it does not match with all materials, it indicates that the pixel point is abnormal.

[0064] S3: Preprocessing the visible light image and the infrared image of the suspected damaged area, and then performing image registration and cropping; and performing statistical analysis to obtain a probability density distribution map of the visible light image and a probability density distribution map of the infrared image.

[0065] The preprocessing method is as follows: denoising, contrast enhancement, and smoothing filtering are performed on the collected visible light image of the suspected damage area; temperature calibration and pseudo-color mapping are performed on the infrared image of the suspected damage area; and the extreme points in the visible light image and the infrared image of the suspected damage area are extracted using the SIFT algorithm as feature points of the image.

[0066] The specific method of extracting feature points of an image using the SIFT algorithm is:

[0067] Step 1: Constructing a scale space. The original image is convolved with Gaussian kernels of varying standard deviations to produce a series of images of varying scales. These images form a pyramidal structure, with each layer increasing in scale and decreasing in resolution.

[0068] Step 2: Detect extreme points. In the constructed scale space, extreme points are detected by comparing the grayscale values ​​of each pixel with its 26 adjacent pixels (including 8 neighboring pixels at the same scale and 9 pixels at the upper and lower adjacent scales). If the grayscale value of a pixel is greater than or less than the grayscale values ​​of all its adjacent pixels, then the pixel is an extreme point. These extreme points are considered representative feature points in the image, and they can maintain relatively stable positions and characteristics at different scales.

[0069] The image registration method is as follows: a rectangular coordinate system is established for each of the pre-processed conventional image and the infrared image, a square consisting of four pixels is defined as a pixel block, and the coordinates (x0, y0) of each pixel block in the pre-processed conventional image rectangular coordinate system and the coordinates (x, y) of each pixel block in the pre-processed infrared image rectangular coordinate system are given; the relative rotation angle is θ, and the relative position translation vector is (a, b). The position formula of the infrared image becomes:

[0070]

[0071] Change the relative rotation angle to θ and the relative position translation vector (a, b) to maximize (x0·x'+y0·y'). At this time, the corresponding relative rotation angle is θ, and the relative position translation vector (a, b) makes the feature points in the infrared image and the traditional image aligned in space.

[0072] S4: Using the registered and cropped visible light image of the suspected damage area, the probability density distribution map of the visible light image, and the infrared image of the suspected damage area, the probability density distribution map of the infrared image as input, respectively, a pre-trained generative adversarial network is used for classification to determine whether the suspected damage area is a damage area.

[0073] Visible light and infrared images of steel components with different types of apparent damage were collected under different lighting and temperature conditions. The two images were cropped to retain the image covering the same area. Image preprocessing methods were then applied to the visible light and infrared images. Denoising, contrast enhancement, and smoothing filtering were performed on the visible light images, while temperature calibration and pseudo-color mapping were performed on the infrared thermal images. Furthermore, image registration was performed to identify the corresponding relationships between the visible light and infrared thermal images, and probability density distributions of the two images were calculated. The damage characteristics under different lighting and temperature conditions were analyzed. Using the two images and their pixel value probability distributions as input and the apparent damage type as output, a generative adversarial network (GAN) was trained to classify the collected damage images and identify different types of damage on the steel surface. Model evaluation and validation were then conducted, and the accuracy of the recognition results was assessed. This included comparison with manual inspection results and calculation of metrics such as false detection rate and missed detection rate, to evaluate the reliability and effectiveness of the method.

[0074] The trained generative adversarial network is used to classify the visible light images and infrared images of suspected damage areas actually collected.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A steel surface damage identification device based on binocular machine vision of visible light images and infrared images, characterized in that: The invention comprises a visible light image capturing device, an infrared image capturing device, a guide rail (1), a controller and a data storage device. The visible light image capturing device and the infrared image capturing device can be slidably arranged on one side of the guide rail (1), the controller and the data storage device are fixed on the other side of the guide rail (1), the visible light image capturing device and the infrared image capturing device are both wirelessly connected to the controller, and the data storage device is connected to the controller.

2. The steel surface damage identification device based on binocular machine vision of visible light images and infrared images according to claim 1 is characterized in that: The visible light image capturing device comprises a first sliding base (3), a conventional camera (5) and a white light source (6), wherein the conventional camera (5) and the white light source (6) are both located on a side of the first sliding base (3) away from the guide rail (1), and the conventional camera (5) is located on a side of the white light source (6).

3. The steel surface damage identification device based on binocular machine vision of visible light images and infrared images according to claim 1 is characterized in that: The infrared image shooting device comprises a second sliding base (11), an infrared camera (8) and an infrared light source (10), wherein the infrared camera (8) and the infrared light source (10) are both located on a side of the second sliding base (11) away from the guide rail (1), and the infrared light source (10) is arranged on one side of the infrared camera (8).

4. The steel surface damage identification device based on binocular machine vision of visible light images and infrared images according to claim 2 is characterized in that: A positioning module I and a positioning module II are respectively provided on the side of the first sliding base (3) close to the guide rail (1) and the side of the second sliding base (11) close to the guide rail (1).

5. The steel surface damage identification device based on binocular machine vision of visible light images and infrared images according to claim 2 or 4, characterized in that: The first sliding base (3) is provided with a wireless receiving module I, a motor driving circuit I and a stepping motor I which are connected in sequence; the second sliding base (11) is provided with a wireless receiving module II, a motor driving circuit II and a stepping motor II which are connected in sequence, and both the wireless receiving module I and the wireless receiving module II are connected to a controller.

6. The steel surface damage identification device based on binocular machine vision of visible light images and infrared images according to claim 5 is characterized in that: The second sliding base (11) is also provided with a high-pressure spray system (9), and the high-pressure spray system (9) is provided on the other side of the infrared camera (8).

7. A method for identifying surface damage of steel based on binocular machine vision of visible light images and infrared images, using the identification device according to claim 1, characterized in that: Including steps: S1: calibrate the surface feature data of the undamaged steel to obtain normal statistical values ​​of visible light image pixel values ​​and normal statistical values ​​of infrared image pixel values ​​of the undamaged steel; S2: Determine the apparent damaged area and the suspected damaged area using a visible light image capturing device and an infrared image capturing device according to the normal statistical values ​​of the visible light image pixel values ​​and the normal statistical values ​​of the infrared image pixel values, and obtain a visible light image of the suspected damaged area and an infrared image of the suspected damaged area; S3: Preprocessing the visible light image and the infrared image of the suspected damaged area, performing image registration and cropping, and performing statistical analysis to obtain a probability density distribution map of the visible light image and a probability density distribution map of the infrared image; S4: Using the registered and cropped visible light image of the suspected damage area, the probability density distribution map of the visible light image, and the infrared image of the suspected damage area, the probability density distribution map of the infrared image as input, respectively, a pre-trained generative adversarial network is used for classification to determine whether the suspected damage area is a damage area and the damage type.

8. The steel surface damage identification method based on binocular machine vision of visible light images and infrared images according to claim 7 is characterized in that: The implementation method of step S2 is: S2.1: The controller controls the traditional camera (5) to move to a position above the surface of the steel material, turns on the white light source (6), illuminates the area A on the surface of the steel material, and uses the traditional camera (5) to collect a visible light image of the steel material. The controller detects whether there is a pixel value in the visible light image that deviates from the normal statistical value of the pixel value of the visible light image. If it deviates, the area A on the surface of the steel material is determined as a suspected damage area, and a visible light image of the suspected damage area is obtained, and step S2.2 is performed. Otherwise, the process moves to the next area on the surface of the steel material for detection; S2.2: After the controller controls the white light source (6) to be turned off, the controller controls the first sliding base (3) to drive the white light source (6) and the traditional camera (5) to move away, and controls the second sliding base (11) to drive the high-pressure spray system (9), the infrared light source (10) and the infrared camera (8) to move to the steel surface area A; S2.3: The controller controls the high-pressure spray system (9) to start, spraying high-pressure water mist to clean the steel surface area A; S2.4: The controller controls the infrared light source (10) to turn on, and after irradiating the steel surface area A for a few seconds, an infrared camera is used to collect an infrared image of the steel surface area A; S2.5: Crop the visible light image and infrared image of area A on the steel surface separately, retain the image of the intersection area, and compare the infrared image of the intersection area. If the deviation from the normal statistical value of the infrared image pixel value exceeds the threshold, it is determined to be an apparent damage area; otherwise, it is still a suspected damage area.

9. The steel surface damage identification method based on binocular machine vision of visible light images and infrared images according to claim 8, characterized in that: The preprocessing method is as follows: denoising, contrast enhancement, and smoothing filtering are performed on the collected visible light image of the suspected damage area; temperature calibration and pseudo-color mapping are performed on the infrared image of the suspected damage area; and the extreme points in the visible light image and the infrared image of the suspected damage area are extracted using the SIFT algorithm as feature points of the image.

10. The steel surface damage identification method based on binocular machine vision of visible light images and infrared images according to claim 9, characterized in that: The image registration method is as follows: establishing rectangular coordinate systems for the preprocessed conventional image and the infrared image respectively, defining a square consisting of four pixels as a pixel block, and giving the coordinates (x0, y0) of each pixel block in the preprocessed conventional image rectangular coordinate system and the coordinates (x, y) of each pixel block in the preprocessed infrared image rectangular coordinate system; The relative rotation angle is θ, the relative position translation vector is (a, b), and the position formula of the infrared image becomes: Change the relative rotation angle to θ and the relative position translation vector (a, b) to maximize (x0·x'+y0·y'). At this time, the corresponding relative rotation angle is θ, and the relative position translation vector (a, b) aligns the feature points in the infrared image and the traditional image in space.

Images