Method and device for detecting and judging cracks of refractory layer of ladle before full-size iron receiving

Abstract

The invention relates to the field of metallurgical equipment state monitoring, and discloses a method and a device for detecting and judging cracks of a refractory layer of a ladle before full-size iron receiving, and the method comprises the steps: obtaining an image of the inner wall of the ladle; performing first-stage processing on the image, and quickly positioning one or more suspected crack regions by using a convolutional neural network; second-stage processing is carried out on each suspected crack area, and three-dimensional physical parameters such as the physical width and the physical depth of the crack in the area are calculated by actively projecting an optical pattern and analyzing geometric deformation of the optical pattern; and carrying out numerical comparison on the calculated physical parameters and a preset safety standard, and carrying out automatic ladle judgment on the ladle. According to the method, through a two-stage processing strategy, the rapid identification capability of deep learning and the accurate quantification capability of active visual three-dimensional measurement are combined, and accurate acquisition of the physical size of the crack is realized, so that reliable data support is provided for automatic and objective packet judgment decision, and the accuracy and reliability of detection are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metallurgical equipment state monitoring, in particular to a full-size ladle refractory layer crack detection and ladle judgment method and device. BACKGROUND

[0002] In the steel smelting process, the ladle is a key container for holding and transporting high-temperature molten iron. The inner wall of the ladle is lined with a refractory layer, which will inevitably produce cracks, spalling and other forms of damage under the cyclic action of high temperature, erosion and chemical corrosion. If these damages are not detected in time and properly treated, it may cause the refractory layer to penetrate, resulting in molten iron leakage and posing a major safety hazard. Therefore, it is essential to detect the health status of the inner wall of the ladle before each use.

[0003] Currently, some automated detection schemes have adopted industrial cameras to capture images of the interior of the ladle, and computer image processing techniques to identify whether there are cracks on the inner wall. Compared with traditional manual visual inspection, this method has improved efficiency and safety. However, most of these methods are limited to two-dimensional image analysis, and the core is to identify the pixel area of suspected cracks based on image brightness, contrast or texture features.

[0004] Such two-dimensional image analysis-based techniques have inherent limitations in that they can answer the questions of "yes or no" and "where" of the cracks, but cannot provide quantitative information about the geometric size of the cracks. That is, it can detect the presence of a crack, but cannot directly measure the actual width or depth of the crack in the three-dimensional physical world. For the safety assessment of the ladle, the physical size of the crack is exactly the most critical indicator to determine whether it can continue to serve safely. Due to the lack of objective and accurate physical size data as support, the final ladle judgment decision often still relies on manual experience to interpret the two-dimensional image, which limits the objectivity, consistency and automation of the entire detection process, making it difficult to achieve fully reliable automated ladle judgment. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application provides a full-size ladle refractory layer crack detection and ladle judgment method and device, which solves the problem that the detection of cracks on the inner wall of the ladle in the prior art relies mainly on two-dimensional image analysis. Although this method can identify the presence of cracks, it cannot obtain physical size information such as actual width or depth, so it lacks objective quantitative basis when making a decision on whether the ladle can continue to be used, resulting in insufficient reliability of the ladle judgment result.

[0006] To achieve the above object, the present application is realized by the following technical scheme:

[0007] The first aspect of the present application provides a method for detecting and judging cracks in the refractory lining of a full-size ladle, comprising the following steps:

[0008] Step one, obtaining an image of the inner wall of the ladle;

[0009] Step two, performing a first-stage processing on the image to quickly locate one or more suspected crack regions using a convolutional neural network;

[0010] Step three, performing a second-stage processing on each suspected crack region to obtain its physical parameters;

[0011] Step four, automatically judging the ladle according to the physical parameters and a preset safety standard.

[0012] In an optional embodiment of the present application, the second-stage processing comprises: controlling a camera module integrated with a controllable light source to project a preset optical pattern onto the suspected crack region; collecting an image superimposed with the optical pattern; and calculating the physical parameters by analyzing the geometric deformation of the optical pattern in the image due to the three-dimensional profile of the surface of the suspected crack region. This two-stage processing method first uses the high efficiency of the convolutional neural network to complete a wide range of fast screening and positioning, and then only targets the local area with potential risks for accurate measurement based on active vision, thereby ensuring the detection efficiency while realizing the quantification of the key physical dimensions of the cracks and providing an objective data basis for subsequent automatic judgment.

[0013] In an optional embodiment of the present application, the physical parameters at least include at least one of the physical width and the physical length of the crack.

[0014] In an optional embodiment of the present application, the optical pattern is a single or multiple parallel laser lines.

[0015] In an optional embodiment of the present application, the step of calculating the physical parameters by analyzing the geometric deformation of the optical pattern specifically comprises: identifying the break or bending of the optical pattern in the image due to the physical depth of the suspected crack region; measuring the pixel distance at the break or bending feature; and converting the pixel distance into the physical parameters in combination with a pre-calibrated pixel-physical size scale. For example, the physical width W of the crack crack can be calculated by the following formula:

[0016] W crack = d pixel · S scale ;

[0017] where d pixel is the measured pixel distance, S scale is the pixel-physical size scale, with the unit of mm / pixel.

[0018] In an optional embodiment of the present application, the first-stage processing further comprises a pre-processing step for the image of the inner wall of the ladle. The pre-processing step comprises at least one of gray scale conversion and image noise reduction, so as to reduce the calculation amount of the subsequent convolutional neural network processing and suppress the interference of image noise.

[0019] In an optional embodiment of the present application, the second-stage processing further comprises a step of parallelly calculating a gradient magnitude image of the suspected crack region. The gradient magnitude image can highlight the regions with sharp gray scale changes in the image, i.e. the edges of the cracks, so as to assist in more accurate identification and measurement of the geometric deformation of the optical pattern. The calculation of the gradient magnitude M grad can be represented by the following formula:

[0020]

[0021] where I x (x,y) and I y (x,y) are the gradients of the image along the horizontal and vertical directions at the pixel point (x,y), respectively.

[0022] In an optional embodiment of the present application, the step of automatically judging the ladle comprises: comparing the maximum crack width in the physical parameters with a preset safety standard threshold value; if the maximum crack width is less than the safety standard threshold value, the ladle is determined to be qualified; if the maximum crack width is greater than or equal to the safety standard threshold value, the ladle is determined to be unqualified.

[0023] In an optional embodiment of the present application, the method further comprises: according to the result of the automatic judgment of the ladle, automatically issuing an instruction to a production scheduling system to perform corresponding production scheduling on the ladle.

[0024] The second aspect of the present application provides a full-size pre-iron ladle refractory layer crack detection and ladle judging device, which comprises:

[0025] an image acquisition unit, which is configured to acquire the image of the inner wall of the ladle;

[0026] A processing unit is electrically connected with the image acquisition unit, and the processing unit is configured to perform the following operations: performing first-stage processing on the image, quickly locating one or more suspected crack regions by using a convolutional neural network; performing second-stage processing on each suspected crack region to obtain physical parameters thereof; and automatically judging the ladle according to the physical parameters and a preset safety standard.

[0027] In the device of the application, the image acquisition unit is a camera module integrated with a controllable light source.

[0028] In the second-stage processing, the processing unit is further configured to control the controllable light source to project a preset optical pattern on the suspected crack region, and calculate the physical parameters by analyzing geometric deformation of the optical pattern based on an image acquired by the image acquisition unit and superimposed with the optical pattern.

[0029] The application provides a full-size pre-ironing ladle refractory layer crack detection and ladle judging method and device.

[0030] The application has the following beneficial effects:

[0031] 1. The application sets the first-stage processing and the second-stage processing, after the suspected crack region is quickly located by using the convolutional neural network, the actual physical parameters of the crack are obtained by projecting the optical pattern and analyzing the geometric deformation, the detection result is improved from the qualitative judgment of "yes or no" on the two-dimensional image level to the quantitative evaluation of "how wide or how long" in the three-dimensional physical space, objective and repeatable data basis is provided for the automatic ladle judging, and subjectivity and uncertainty caused by relying on manual experience or pure two-dimensional image analysis are avoided.

[0032] 2. The two-stage cooperative processing mechanism adopted in the application firstly performs global and rapid screening on the inner wall of the ladle by using the convolutional neural network with high processing speed, and quickly locks the potential problem region, then only the suspected region in a small range is targeted for more accurate but relatively time-consuming geometric quantization, the strategy avoids time-consuming three-dimensional scanning on the entire inner wall of the ladle, optimizes the allocation of computing resources, greatly shortens the period of single detection without sacrificing the measurement accuracy of the key region, and meets the fast-paced demand of the industrial production line.

[0033] 3、The present application integrates the positioning, quantification and final decision of cracks in a complete automated process. By automatically comparing the quantified physical parameters with the preset safety standards, the system can directly output a clear conclusion of "qualified" or "unqualified", and can further automatically send instructions to the production scheduling system. This not only reduces the workload of manual review, reduces the possibility of misjudgment caused by personnel fatigue or standard execution, but also makes the health status management of the ladle more standardized and efficient. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a structural schematic diagram of the present application;

[0035] Figure 2 is an image preprocessing schematic diagram of the present application;

[0036] Figure 3 is a ladle decision flowchart of the present application;

[0037] Figure 4 is a schematic diagram for calculating the crack depth using the principle of structured light triangulation of the present application.

[0038] Wherein, 1, track; 2, empty ladle; 3, detection room; 4, camera; 5, computer; 6, trolley; 7, portal track; 8, carrier. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0040] Referring to the drawings Figure 1 , Figure 1 is a structural schematic diagram of the full-size ladle refractory crack detection and ladle decision device before ironing according to an embodiment of the present application. The device is deployed in a detection room 3 for detecting an empty ladle 2. The empty ladle 2 is transported to the predetermined detection position in the detection room 3 by a carrier 8 on a track 1.

[0041] The device comprises a portal track 7 which spans above the position of the aerial ladle 2. A hoist trolley 6 is arranged on the portal track 7 and can move horizontally along the track. A multi-modal sensor head 4 is suspended and fixed on the hoist trolley 6 and can move vertically through the lifting mechanism of the hoist trolley 6. Through the horizontal movement of the portal track 7 and the vertical movement of the hoist trolley 6, the multi-modal sensor head 4 can enter the interior of the aerial ladle 2 and reach any position of the inner wall for data collection.

[0042] A camera for acquiring images of the inner wall of the aerial ladle 2 and a controllable light source for projecting a preset optical pattern on the inner wall of the aerial ladle 2 at a specific processing stage are integrated inside the multi-modal sensor head 4. The multi-modal sensor head 4 is electrically connected to a computer 5 through a line 9. The computer 5 is used to receive and process image data collected by the multi-modal sensor head 4 and execute the detection and ladle judgment method of the present application.

[0043] Referring to the accompanying drawings Figure 2 , Figure 2 is a method flowchart of the full-size pre-ironing ladle refractory layer crack detection and ladle judgment method according to an embodiment of the present application. The present application provides a full-size pre-ironing ladle refractory layer crack detection and ladle judgment method, which is executed by a computer 5 and can include the following steps:

[0044] Step S101, acquire images of the inner wall of the ladle. Specifically, after the aerial ladle 2 is transported to the detection position, the computer 5 controls the portal track 7 and the hoist trolley 6 to drive the multi-modal sensor head 4 to descend to a preset starting height inside the aerial ladle 2. Subsequently, the multi-modal sensor head 4 scans the entire inner wall of the aerial ladle 2 along a predetermined trajectory (e.g., rotates and ascends), and its built-in camera continuously collects images during the process and transmits image signals to the computer 5 in real time through the line 9.

[0045] Step S102, perform first-stage processing on the images to quickly locate one or more suspected crack regions using a convolutional neural network. A processing program is deployed inside the computer 5, which includes a first-stage processing module a. The module a receives images from step S101 and analyzes the images using a pre-trained convolutional neural network model to identify regions in the images with crack visual features and outputs the position coordinate information of these regions, forming a set of suspected crack regions.

[0046] Step S103, a second stage processing is performed on each suspected crack region to obtain its physical parameters. The program of computer 5 further comprises a second stage processing module b. For each suspected crack region outputted in step S102, this module b performs a targeted processing. It first sends a command to multi-modal sensor head 4 through line 9 to control its internal controllable light source to project a preset optical pattern to this specific region. Meanwhile, camera captures an image with the optical pattern superimposed. Subsequently, module b analyzes the geometric deformation of the optical pattern on this image and calculates the physical parameters of this suspected crack region, such as physical width, based on the analysis result.

[0047] Step S104, an automatic ladle judgment is performed on the ladle according to the physical parameters and a preset safety standard. The program of computer 5 further comprises an automatic ladle judgment module c. This module c compares the physical parameters of all cracks calculated in step S103 with a safety standard stored locally or called from a database. Based on the comparison result, module c generates a clear “qualified” or “unqualified” judgment conclusion.

[0048] After the judgment is completed, the method can further comprise a subsequent step: according to the judgment conclusion generated by automatic ladle judgment module c, computer 5 sends corresponding scheduling instructions to a production scheduling system (not shown in the figure) through line 9. For example, if the judgment conclusion is “qualified”, it notifies the tapping platform that this ladle can be used for tapping; if the conclusion is “unqualified”, it notifies the scheduling room to send this ladle to a repair station.

[0049] In a specific embodiment, after receiving the original image captured by multi-modal sensor head 4, the method first performs an image preprocessing step. The purpose of this step is to reduce the computational complexity of subsequent processing and suppress the noise introduced in the image capturing process. The preprocessing can include the following operations:

[0050] Step one, grayscale conversion. Convert the multi-channel color RGB image into a single-channel grayscale image I gray . This conversion is achieved by weighted summation of intensity values of each color channel, and its calculation process can be represented as:

[0051] I gray (x,y)=W R ·C R (x,y)+W G ·C G (x,y)+W B ·C B (x,y);

[0052] In the formula, I gray (x,y) is the output grayscale value at pixel coordinate (x,y); C R (x,y), CG (x,y),C B (x,y) respectively are the original intensity values of red, green and blue channels of the pixel point; W R′ W G′ W B are preset and fixed weight coefficients.

[0053] Step two, image denoising. In order to reduce the influence of random noise in the image on the subsequent crack identification, a Gaussian filter can be used to smooth the gray-scale image I gray , and a denoised image I gauss is obtained. This process is realized by a convolution operation between a Gaussian kernel and the image. After the pre-processing is completed, the first stage processing module a processes the denoised image I gauss . A pre-trained convolutional neural network is deployed in this module a. Through a series of convolutional layers and pooling layers, this network can extract hierarchical features from the input image, such as edge, texture and shape combination features. Through the learning and analysis of these features, the network can identify the areas in the image that meet the crack morphological features. Finally, the module a outputs the position information of one or more suspected crack areas, for example, giving the coordinate range of each suspected area in the image in the form of a bounding box.

[0054] Subsequently, the second stage processing module b receives the set of suspected crack areas output from the first stage processing module a, and performs targeted geometric quantification processing on each area in the set. Referring to the accompanying drawings, Figure 3 , Figure 3 is a schematic diagram of the principle of the second stage processing according to an embodiment of the present application. The processing can include:

[0055] Step S201, projecting an optical pattern. The second stage processing module b sends instructions to the multi-modal sensing head 4 through the line 9 to control the controllable light source inside it, and only projects one or more laser lines with known geometric patterns to the suspected crack area being processed, forming an optical pattern. In an embodiment, the optical pattern can be a set of parallel fine laser lines.

[0056] Step S202, collecting a deformation image. The camera built-in the multi-modal sensing head 4 collects the image of the suspected crack area illuminated by the optical pattern. Since the crack itself is a three-dimensional structure with physical depth, when the flat laser line is projected onto its surface, it will appear discontinuous fracture or local bending on the two-dimensional collected image, and this phenomenon is called geometric deformation.

[0057] Step S203: Calculate physical parameters. The second-stage processing module b analyzes the image acquired in step S202. It first automatically identifies breakpoints or bending feature points on the optical pattern using image processing algorithms, and measures the pixel distance d of these feature points in the image coordinate system. pixel Subsequently, module b utilizes a pixel-to-physical size scale S pre-determined during the system calibration phase. scale The measured pixel distance is converted into the actual physical width W of the crack. crack Its calculation can be expressed by the following formula:

[0058] W crack =d pixel ·S scale ;

[0059] By measuring multiple locations on the optical pattern, the width of the crack at different positions can be obtained, and its maximum width can be determined. In an optional implementation, to aid in the accurate identification of geometric deformation, the second-stage processing module b can also calculate the gradient magnitude image M of the suspected crack region in parallel. grad Regions with high gradient magnitudes correspond to edges in the image where grayscale changes drastically, clearly outlining the contours of cracks. The gradient magnitude can be calculated as follows:

[0060]

[0061] In the formula, I x (x,y) and I y (x, y) represent the first derivatives of the original image at pixel (x, y) along the horizontal and vertical directions, respectively, which can be obtained by convolving the image with gradient operators such as the Sobel operator. After completing the second-stage processing of all suspected crack regions, the automated crack detection module c summarizes the physical parameters of all confirmed and quantized cracks. This module c compares the maximum value of the maximum physical width of all cracks with a configurable safety standard threshold T stored in computer 5 or retrieved from an external database. width Compare the values. If the maximum width value is less than the safety standard threshold T... width If the maximum width value is greater than or equal to the safety standard threshold T, then the package is judged as "qualified"; width If the result is not satisfactory, the ladle will be judged as "unqualified". This conclusion is the final automated judgment result for that ladle.

[0062] In one specific embodiment, after receiving the raw image acquired by the multimodal sensor head 4, the method first performs an image preprocessing step. The purpose of this step is to reduce the computational complexity of subsequent processing and suppress noise introduced during image acquisition. Preprocessing may include the following operations:

[0063] Step one, gray scale conversion. Convert the multi-channel color RGB image into a single-channel gray scale image I gray . The conversion is achieved by weighted summation of intensity values of each color channel, whose calculation process can be represented as:

[0064] I gray (x, y) = R(x, y) I R + G(x, y) I G + B(x, y) I B ;

[0065] where I gray (x, y) is the output gray scale value at pixel coordinate (x, y); R(x, y), G(x, y), B(x, y) are the red, green, blue color levels of the pixel, respectively, whose values are integers from 0 to 255; I R , I G , I B are the gray scale coefficients of red, green, blue color levels, respectively, in one embodiment, I R = 0.31, I G = 0.60, and I B = 0.11. Step two, image denoising. To reduce the impact of random noise in the image on subsequent crack identification, a Gaussian filter can be used to smooth the gray scale image I gray , obtaining the denoised image I gauss . The process is achieved by a convolution operation between a Gaussian kernel and the image, whose mathematical model is as follows:

[0066]

[0067] where G kernel (i, j) is the value of the Gaussian kernel function at coordinate (i, j), and k is the size of the Gaussian kernel. After completing the preprocessing, the first stage processing module a processes the denoised image I gauss . The module a is internally deployed with a pre-trained convolutional neural network. The network extracts hierarchical features from the input image through a series of convolutional and pooling layers to identify regions in the image that meet the morphological characteristics of cracks. Finally, the module a outputs the position information of one or more suspected crack regions.

[0068] Subsequently, the second stage processing module b receives the set of suspected crack regions output from the first stage processing module a and performs targeted geometric quantification processing on each region in the set. Referring to Appendices Figure 3 and Appendices Figure 4 , the processing can include:

[0069] Step S201, project optical pattern and collect deformation image. The second stage processing module b controls the controllable light source inside the multi-modal sensor head 4 to project a set of parallel laser lines to the suspected crack area, and collects the image superimposed with the optical pattern by the camera. Due to the physical depth and width of the crack, the straight laser lines will show geometric deformation such as breaking or bending on the two-dimensional image.

[0070] Step S202, calculate physical parameters. The second stage processing module b analyzes the collected deformation image to quantify the physical size of the crack.

[0071] For the physical width W of the crack crack , the module b measures the lateral pixel distance d pixel between the breaking of the laser lines on the edges of the crack on the image coordinate system, and converts it using the pre-calibrated scale S scale :

[0072] W crack = d pixel · S scale ;

[0073] For the physical depth h of the crack, referring to the triangulation principle shown in Figure 4 , the calculation method is as follows. The module b identifies the pixel displacement d shift of the laser lines on the image due to the change in depth, and calculates according to the known system geometric parameters:

[0074]

[0075] In the formula, d shift is the observed pixel displacement; S scale is the pixel-physical size scale; a is the laser projection angle; b is the camera viewing angle.

[0076] In an optional embodiment, to assist in accurate identification of geometric deformation, the second stage processing module b can also calculate the gradient amplitude image of the suspected crack area image in parallel. This calculation first obtains the gradients I x and I y in the x and y directions, which is calculated by convolving the image I (e.g. I gauss ) with the gradient operator (such as the Sobel operator) G x and G y :

[0077]

[0078] Subsequently, the total gradient amplitude M grad (x,y) at the pixel point (x,y) is calculated according to the obtained directional gradient to enhance the crack boundary:

[0079]

[0080] After the second stage processing is completed for all suspected crack areas, the automated ladle classification module c compiles all the quantified physical parameters of the cracks. The module c compares the maximum physical width and the maximum physical depth of all the cracks with a preset safety standard threshold. If all the quantified parameters are less than the corresponding safety standard threshold, the classification result is "qualified"; if any parameter is greater than or equal to the corresponding safety standard threshold, the classification result is "unqualified". This result is the final automated ladle classification result for the ladle.

[0081] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications, changes, omissions, substitutions and adaptations can be made by those skilled in the art without departing from the application, which is defined by the following claims and their equivalents.

Claims

1. A method for detecting and judging cracks in a full-size ladle refractory lining before iron feeding, characterized in that, The method comprises: Step 1: obtaining an image of the inner wall of the ladle; Step 2: performing first-stage processing on the image to quickly locate one or more suspected crack regions by using a convolutional neural network; Step 3: performing second-stage processing on each suspected crack region to obtain physical parameters thereof; Step 4: automatically judging the ladle according to the physical parameters and a preset safety standard.

2. The full size ladle refractory lining crack detection and ladle acceptance method as claimed in claim 1, wherein, The second-stage processing comprises: controlling a camera module integrated with a controllable light source to project a preset optical pattern on the suspected crack region; collecting an image superimposed with the optical pattern; calculating the physical parameters by analyzing the geometric deformation of the optical pattern on the suspected crack region.

3. The full size tundish refractory lining crack detection and tundish identification method according to claim 1, characterized in that, The physical parameters at least include at least one of a physical width and a physical length of the crack.

4. The full size ladle refractory lining crack detection and ladling method according to claim 2, characterized in that, The optical pattern is a single or multiple parallel laser lines.

5. The full size ladle refractory lining crack detection and ladling method according to claim 2, wherein, The step of calculating the physical parameters by analyzing the geometric deformation of the optical pattern on the suspected crack region specifically comprises: identifying a break or a bend of the optical pattern caused by the physical depth of the suspected crack region; measuring a pixel distance at the break or the bend; converting the pixel distance into the physical parameters in combination with a pre-labeled pixel-physical size scale.

6. The full size ladle refractory lining crack detection and ladling method as claimed in claim 1, wherein, Before the first-stage processing, the method further comprises a step of pre-processing the image of the inner wall of the ladle; The step of pre-processing comprises at least one of gray scale conversion and image noise reduction.

7. The full size tundish refractory lining crack detection and tundish identification method according to claim 1, characterized in that, In the second-stage processing, the method further comprises calculating a gradient amplitude image of the suspected crack region in parallel to enhance the crack boundary and assist the calculation of the physical parameters.

8. The full size ladle refractory lining crack detection and ladling method as claimed in claim 1, wherein, The step of automatically judging the ladle specifically comprises: comparing a maximum crack width in the physical parameters with a preset safety standard threshold; if the maximum crack width is less than the safety standard threshold, determining that the ladle is qualified; if the maximum crack width is greater than or equal to the safety standard threshold, determining that the ladle is unqualified.

9. The full size ladle refractory lining crack detection and ladling method according to claim 1, wherein, The method further comprises: automatically sending an instruction to a production scheduling system according to the result of the automatic judgment of the ladle to perform corresponding production scheduling on the ladle.

10. A full-size ladle refractory crack detection and ladle judging device for performing the full-size ladle refractory crack detection and ladle judging method according to any one of claims 1 to 9, characterized by, The method comprises: an image collection unit configured to obtain an image of the inner wall of the ladle; a processing unit electrically connected to the image collection unit, the processing unit being configured to: perform first-stage processing on the image to quickly locate one or more suspected crack regions by using a convolutional neural network; perform second-stage processing on each suspected crack region to obtain physical parameters thereof; automatically judge the ladle according to the physical parameters and a preset safety standard; wherein the image collection unit is a camera module integrated with a controllable light source; and in the second-stage processing, the processing unit is further configured to control the controllable light source to project a preset optical pattern on the suspected crack region, and calculate the physical parameters by analyzing the geometric deformation of the optical pattern based on the collected image superimposed with the optical pattern.

Images