High-temperature plate blank image processing method and system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent image processing, in particular to a high-temperature plate blank image processing method and system based on artificial intelligence. The method comprises the following steps: acquiring an original high-temperature image, dynamically compressing a highlight area, then separating an illumination component and a reflection component, and enhancing the contrast of the reflection component to obtain a preprocessed image; denoising the preprocessed image in combination with guiding filtering and a wavelet threshold to obtain a first processed image; simulating a heat wave distortion effect through elastic transformation of the first processed image to obtain a second processed image, and applying random brightness disturbance to the second processed image to obtain a third processed image; constructing a lightweight model to generate a tiny defect mask, and performing random transformation on the third processing image to obtain a composite image; and based on the composite image and the defect mask, generating a high-temperature image data set and carrying out data analysis to obtain defect features. According to the invention, the problem that a traditional enhancement method lacks high-temperature scene authenticity is solved, and the feature loss rate in an image processing chain is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent image processing, and particularly to an image processing method and system for high-temperature slab based on artificial intelligence. Background Art

[0002] In the industrial production of high-temperature slabs, real-time and accurate surface defect detection is a key link to ensure product quality. However, in a high-temperature environment, usually reaching 800 - 1200 °C, it causes a strong thermal radiation effect on the surface of the slab, specifically manifested as an extreme dynamic range of the image, such as the coexistence of local overexposure and dark areas, non-uniform brightness interference, such as fluctuations in the reflection of scale and dense thermal noise, and disturbances of heat wave airflows. Traditional image processing methods face significant challenges: The dynamic range compression algorithm with fixed parameters cannot adaptively adjust the highlight suppression intensity, resulting in the loss of texture details of tiny cracks or pits; The filtering and denoising model designed based on artificial experience can suppress random noise, but it is easy to blur the edge features of defects, causing feature confusion; In addition, existing data augmentation techniques mostly rely on geometric transformations and lack the simulation of high-temperature physical characteristics, resulting in insufficient generalization ability of the model. The information loss in such an image processing chain directly affects the defect recognition accuracy. According to statistics, the false detection rate of traditional methods in high-temperature scenarios is as high as 30% - 40%, seriously restricting the yield rate and automation level of the production line. Therefore, how to reduce the feature loss rate in the image acquisition and preprocessing stage has become the core breakthrough point for improving the accuracy of high-temperature industrial quality inspection. Summary of the Invention

[0003] The object of the present invention is to propose an image processing method and system for high-temperature slab based on artificial intelligence in view of the problems existing in the background art.

[0004] The technical solution of the present invention: An image processing method for high-temperature slab based on artificial intelligence includes the following steps: S1. Obtain the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain a preprocessed image; Finally, denoise the preprocessed image by combining guided filtering and wavelet thresholding to obtain a first processed image; S2. Simulate the heat wave distortion effect on the first processed image through elastic transformation to obtain a second processed image, and apply random brightness perturbation to the second processed image to obtain a third processed image; S3. Obtain the normal slab image, the noise vector of the original high-temperature image, and the slab defect type, construct a lightweight model to generate a tiny defect mask, and perform random transformation on the third processed image based on the defect mask to obtain a synthetic image; S4. Generate a high-temperature image dataset based on the synthetic image and the defect mask, and perform data analysis to obtain the defect features of the high-temperature original image.

[0005] Preferably, for S1, obtain the original high-temperature image , where H and W are the pixel height and width respectively; the method for dynamically compressing the highlight area of the original high-temperature image includes: Obtain the initial brightness value of each pixel point of the original high-temperature image, and perform dynamic compression on the original high-temperature image based on the initial brightness value. The compression formula is as follows: ; In the formula, is the processed brightness value; I(i, j) is the initial brightness value; D is the maximum value of the standard brightness range; α1 is an adaptive parameter, calculated according to the dynamic range of the image, , μ1 is the mean value of the brightness values of the original high-temperature image, and σ1 is the variance of the brightness values; Based on the processed brightness value of each pixel point , generate a preprocessed image.

[0006] Preferably, the method for analyzing and obtaining the separated illumination component and reflection component and enhancing the contrast of the reflection component includes: Perform convolution on the preprocessed image using three Gaussian kernels. The convolution formula is as follows: ; In the formula, is the illumination component based on the Gaussian kernel; σ2 is the smoothing parameter used to control the Gaussian kernel, and σ2 ∈ {15, 80, 200}.

[0007] Preferably, take the logarithm of the illumination component at each scale and subtract them to obtain the corresponding reflection component at multiple scales , and the calculation formula is as follows: ; The final reflection component R(x) is calculated by fusing and averaging the multi-scale reflection components. The formula is as follows: .

[0008] Preferably, perform enhancement processing on the final reflection component R(x), stretch the histogram of the final reflection component R(x) to [0, D], and apply CLAHE (Contrast Limited Adaptive Histogram Equalization); The method for combining guided filtering and wavelet threshold denoising to obtain the first processed image includes: Use the final reflection component R(x) as the guidance map, filter the original high-temperature image, perform 3-layer wavelet decomposition on the filtered image to obtain high-frequency subbands and low-frequency subbands, and apply a global hard threshold to the high-frequency subbands to obtain the first processed image.

[0009] Preferably, for S2, the method of simulating the heatwave distortion effect through elastic transformation to obtain the second processed image includes: Create two independent Gaussian filter displacement matrices and , the size of the Gaussian filter displacement matrix is the same as that of the first processed image, and the two Gaussian filter displacement matrices and are calculated and obtained based on the following formula: , ; In the formula, α2 is the displacement amplitude; σ4 is the Gaussian kernel standard deviation; is a two-dimensional Gaussian distribution random number; i and j are the horizontal and vertical coordinates of the pixel points in the processed image respectively; Calculate the new coordinates for each pixel point (i, j) of the processed image, and the calculation formula is as follows: , ; Update the coordinates of each pixel point according to the new coordinates, and resample the processed image after calculating the new coordinates using bilinear interpolation to obtain the second processed image, avoiding the jagged effect.

[0010] Preferably, the method of applying random brightness perturbation to the second processed image to obtain the third processed image includes: Extract the brightness channel of the processed image, and perform random scaling on the brightness channel. The random scaling expression is as follows: ; In the formula, is the scaled brightness; V is the original brightness; δ is the random scaling parameter, δ ∈ (-0.3, 0.3); The method of constructing a lightweight model to generate a micro defect mask includes: Define the defect type label using one-hot encoding, input the noise vector and the defect type label into the defect generator network to obtain the defect mask M, and perform normalization processing on the defect mask M; Superimpose the defect mask M on the normal sheet image at a random position and a random rotation angle for transformation and scaling to obtain the transformed image. The transformation and scaling are implemented based on the following formula: ; In the formula, is the synthesized image; is the normal sheet image; is the defect image; the symbol "⊙" represents element-wise multiplication.

[0011] Preferably, the method for learning the defect features of the high-temperature original image includes: Load a batch of synthetic images and defect masks M from the high-temperature image dataset, and input the preprocessed MobileNetV3-small model with the synthetic images of the same batch to obtain the original output result. Normalize the original output result to the probability space through the Sigmoid activation function to obtain the defect probability. Calculate the classification loss for each pixel point (i, j), and the calculation formula is as follows: ; In the formula, z(i, j, b) is the true mask label, z(i, j, b) ∈ {0, 1}; P(i, j, b) is the defect probability of the pixel point; b is the synthetic image number of the same batch, b is a positive integer, b ∈ [1, B], B is the size of the β batch; is the positive sample weight; F() is the loss function.

[0012] Preferably, the expression of the loss function is as follows: ; In the formula, a, b, and c are all variables; Sum and normalize all pixels through the following formula: ; In the formula, is the total loss; Take the total loss as the defect feature of the original high-temperature image.

[0013] The present invention also discloses an artificial intelligence-based high-temperature slab image processing system, which applies the above-mentioned artificial intelligence-based high-temperature slab image processing method, and specifically includes: An image acquisition and first processing module, used to acquire the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain a preprocessed image; finally, denoise the preprocessed image by combining guided filtering and wavelet threshold to obtain a first processed image; A second processing module, used to simulate the heat wave distortion effect on the first processed image through elastic transformation to obtain a second processed image, and apply random brightness perturbation to the second processed image to obtain a third processed image; A third processing module, used to acquire the normal plate image, the noise vector of the original high-temperature image, and the plate defect type, construct a lightweight model to generate a micro defect mask, and perform random transformation on the third processed image based on the defect mask to obtain a synthetic image; An image analysis and fourth processing module, based on the synthetic image and the defect mask, generates a high-temperature image dataset and performs data analysis to obtain the defect features of the high-temperature original image.

[0014] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: (1) Through dynamic logarithmic tone mapping, the compression intensity is adaptively adjusted based on the image mean and variance, avoiding local information loss caused by fixed parameters; (2) Adopt the fusion of multi-scale decomposition and CLAHE, separate the non-uniform illumination component through a three-scale Gaussian kernel, and enhance the reflection characteristics of minute defects by combining block histogram equalization; Design guided filtering-wavelet threshold joint denoising to suppress thermal radiation noise while retaining edges; through high-temperature data augmentation, innovatively introduce physically constrained defect synthesis and elastic heatwave distortion, solve the problem of lack of authenticity in high-temperature scenes in traditional enhancement methods, reduce the feature loss rate in the image processing chain, and provide high-fidelity input for subsequent analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a method block diagram of Embodiment 1 proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Embodiment 1, as Figure 1 shown, the artificial intelligence-based high-temperature slab image processing method proposed by the present invention includes the following steps: S1. Obtain the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain a preprocessed image; finally, denoise the preprocessed image by combining guided filtering and wavelet threshold to obtain a first processed image; For S1, obtain the original high-temperature image , where H and W are the pixel height and width respectively; obtain the initial brightness value of each pixel point of the original high-temperature image, and perform dynamic compression on the original high-temperature image based on the initial brightness value to obtain a preprocessed image. The compression formula is as follows: ; In the formula, is the processed brightness value; I(i, j) is the initial brightness value; D is the maximum value of the standard brightness range; α1 is an adaptive parameter calculated according to the image dynamic range, , μ1 is the mean value of the brightness values of the original high-temperature image, and σ1 is the variance of the brightness values; Based on the processed brightness value of each pixel point, generate a preprocessed image; Due to the logarithmic characteristic of the compression formula, the dark part gradient is amplified while the highlight gradient is compressed, making the overall image more adaptable to the perception characteristics of the human eye or algorithms; Exemplarily, areas that were originally overexposed and whitened due to high temperature are non-linearly compressed, such as hot spots on the surface of the sheet, revealing textures and structures, such as oxidation marks or minute protrusions. The brightness values are mapped from the extremely high dynamic range of the original high-temperature image to a standard range. For example, a brightness level of 0 to 10,000 is mapped to the standard 8-bit range (0 to 255), avoiding detail loss caused by sensor saturation; Also, for example, defects originally hidden in the dark areas are significantly enhanced, and the edge sharpness is improved, such as cracks, pits, etc.; Methods for analyzing and obtaining separated illumination components and reflection components and enhancing the contrast of the reflection components include: Convolve the preprocessed image using three Gaussian kernels. The convolution formula is as follows: ; In the formula, is the illumination component based on the Gaussian kernel; σ2 is the smoothing parameter used to control the Gaussian kernel, and σ2 ∈ {15, 80, 200}; among them, when σ2 = 15, it is used to retain local brightness changes; when σ2 = 200, it is used to capture global illumination; when σ2 = 80, it is used to balance local brightness changes and global illumination; Take the logarithm of the illumination component at each scale and subtract them to obtain the corresponding reflection components at multiple scales , and the calculation formula is as follows: ; The final reflection component R(x) is calculated by fusing and averaging the multi-scale reflection components. The formula is as follows: ; Perform enhancement processing on the final reflection component R(x). Stretch the histogram of the final reflection component R(x) to [0, D], and apply CLAHE (Contrast Limited Adaptive Histogram Equalization); It should be noted that the maximum value D of the standard brightness range is taken as 255, so the standard brightness range is [0, 255]. Stretching the histogram of the final reflection component R(x) to [0, 255] is implemented based on the following formula: ; In the formula, and are the maximum value and minimum value of the final reflection component respectively; The method of applying CLAHE (Contrast Limited Adaptive Histogram Equalization) is as follows: Divide the preprocessed image into 8×8 small blocks, perform histogram equalization on each small block separately to enhance local contrast, set Clip Limit = 2.0 to prevent noise amplification; if the histogram of a certain block exceeds the threshold, evenly distribute the excess part to other brightness levels, and smooth the block boundaries through interpolation to avoid block effects; A method for obtaining a first processed image by combining guided filtering and wavelet threshold denoising includes: Using the final reflection component R(x) as a guidance map, filtering the original high-temperature image, performing three-layer wavelet decomposition on the filtered image to obtain high-frequency subbands and low-frequency subbands, and applying a global hard threshold to the high-frequency subbands to obtain a first processed image; It should be noted that the method for filtering the original high-temperature image includes setting a filtering radius r, for example, the filtering radius r = 5, and a regularization coefficient = 0.01; setting that the filtering output result is a linear transformation of the guidance map G within a local window Then, the filtering output result is obtained based on the following formula: ; In the formula, is the filtering output result, h is the filtering output result number, and h is a positive integer; k is the local window number, and k is a positive integer; Solving the coefficients and by minimizing the cost function, and the cost function to be minimized is as follows: ; Taking the average of the filtering output results of the same local window to obtain a first denoised image; removing large-scale non-uniform noise to provide a more "clean" input for wavelet denoising and reducing the noise energy in the high-frequency subbands; A method for performing three-layer wavelet decomposition on the filtered image to obtain high-frequency subbands and low-frequency subbands, and applying a preset hard threshold to the high-frequency subbands to obtain a first processed image includes: Using the Daubechies-8 wavelet basis to decompose the first denoised image into three layers to obtain low-frequency approximation coefficients (LL3) and high-frequency detail coefficients (LH3, HL3, HH3, LH2, HL2, HH2, LH1, HL1, HH1); Estimating the noise standard deviation σ3 based on the high-frequency subband (HH1), and the calculation formula is as follows: ; Setting a global hard threshold T = 1.5σ3, applying the global hard threshold T to the high-frequency subbands (LH, HL, HH), and the application of the global hard threshold T is implemented based on the following expression: ; Performing inverse wavelet transform on the processed coefficients to obtain the final denoised image; precisely hitting the residual high-frequency noise and avoiding the loss of details caused by global filtering; S2. Simulate the heat wave distortion effect by performing elastic transformation on the first processed image to obtain a second processed image, and apply random brightness perturbation to the second processed image to obtain a third processed image; Regarding S2, the method for simulating the heat wave distortion effect by elastic transformation to obtain the second processed image includes: Create two independent Gaussian filter displacement matrices and , the size of the Gaussian filter displacement matrix is the same as that of the first processed image, and the two Gaussian filter displacement matrices and are calculated and obtained based on the following formula: , ; In the formula, α2 is the displacement amplitude; σ4 is the Gaussian kernel standard deviation; is a two-dimensional Gaussian distribution random number; i and j are the horizontal and vertical coordinates of the pixel points in the processed image respectively; Calculate the new coordinates for each pixel point (i, j) of the processed image, and the calculation formula is as follows: , ; Update the coordinates of each pixel point according to the new coordinates, and resample the processed image after calculating the new coordinates using bilinear interpolation to obtain the second processed image, avoiding the aliasing effect; The method for applying random brightness perturbation to the second processed image to obtain the third processed image includes: Extract the brightness channel of the processed image, and perform random scaling on the brightness channel. The random scaling expression is as follows: ; In the formula, is the scaled brightness; V is the original brightness; δ is the random scaling parameter, δ ∈ (-0.3, 0.3); it should be noted that (-0.3, 0.3) represents a uniformly distributed random number, which is used to limit the brightness change range to ±30%; S3. Obtain the normal sheet image, the noise vector of the original high-temperature image, and the sheet defect type, construct a lightweight model to generate a micro-defect mask, and perform random transformation on the third processed image based on the defect mask to obtain a synthetic image; The method for constructing a lightweight model to generate a micro-defect mask includes: Define the defect type label using One-hot encoding, input the noise vector and the defect type label into the defect generator network to obtain the defect mask M, and perform normalization processing on the defect mask M; The defective mask M is superimposed on the normal plate image at random positions and random rotation angles for transformation and scaling to obtain a transformed image. The transformation and scaling are implemented based on the following formula: ; In the formula, is the synthesized image; is the normal plate image; is the defective image; the symbol "⊙" represents element-wise multiplication; it should be noted that the synthesized image , the normal plate image and the defective image are all of the same size; Exemplarily, the normal plate image ; the defective image ; the defective mask ; then the calculation of the formula includes the following process: ; ; ; The result represented by indicates the defective area: the pixel values in the upper right and lower left areas are reduced to 50, and the normal area remains 200; S4. Based on the synthesized image and the defective mask, generate a high-temperature image dataset and perform data analysis to obtain the defective features of the high-temperature original image; The method for learning the defective features of the high-temperature original image includes: Load a batch of synthesized images and the defective mask M from the high-temperature image dataset, and input them into the pre-processed MobileNetV3-small model of the same batch of synthesized images to obtain the original output result. Normalize the original output result to the probability space through the Sigmoid activation function to obtain the defective probability; Calculate the classification loss for each pixel point (i, j). The calculation formula is as follows: ; In the formula, z(i, j, b) is the true mask label, z(i, j, b) ∈ {0, 1}; P(i, j, b) is the defective probability of the pixel point; b is the number of the synthesized images in the same batch, b is a positive integer, b ∈ [1, B], B is the size of the β batch; is the positive sample weight; F() is the loss function; Preferably, the expression of the loss function is as follows: ; In the formula, a, b, and c are all variables; Sum and normalize all pixels through the following formula: ; In the formula, is the total loss; Take the total loss as the defect feature of the original high-temperature image; Exemplarily, after the input image is processed by the backbone network MobileNetV3-small, the final size of the feature map obtained is B×C×H×W, where C is the number of channels, which is set to 1 in the defect detection task. At the end of the backbone network, the number of channels is compressed to 1 through a 1×1 convolutional layer to obtain the original output result Logits. The original output result Logits satisfies the following format: ; It should be noted that the Sigmoid activation function is an existing function technology and will not be elaborated here; The advantages of the present invention include: (1) Through dynamic logarithmic tone mapping, the compression intensity is adaptively adjusted based on the image mean and variance to avoid local information loss caused by fixed parameters; (3) Adopt multi-scale Retinex decomposition and CLAHE fusion, separate the non-uniform illumination component through a three-scale Gaussian kernel, and enhance the reflection characteristics of micro-defects by combining block histogram equalization; (4) Design guided filtering-wavelet threshold joint denoising to suppress thermal radiation noise while retaining edges; Through high-temperature data augmentation, defect synthesis with physical constraints and elastic heatwave distortion are innovatively introduced, solving the problem of the lack of authenticity of high-temperature scenes in traditional enhancement methods, reducing the feature loss rate in the image processing chain, and providing a high-fidelity input for subsequent analysis.

[0017] Embodiment 2, the high-temperature slab image processing system based on artificial intelligence proposed by the present invention is applied to the high-temperature slab image processing method based on artificial intelligence proposed in Embodiment 1, and specifically includes: An image acquisition and first processing module, which is used to acquire the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain a preprocessed image; Finally, denoise the preprocessed image by combining guided filtering and wavelet threshold to obtain a first processed image; A second processing module, which is used to simulate the heatwave distortion effect on the first processed image through elastic transformation to obtain a second processed image, and apply random brightness perturbation to the second processed image to obtain a third processed image; A third processing module, which is used to obtain a normal sheet image, the noise vector of the original high-temperature image, and the sheet defect type, construct a lightweight model to generate a micro-defect mask, and perform random transformation on the third processed image based on the defect mask to obtain a synthetic image; The image analysis and fourth processing module generates a high-temperature image dataset based on the synthetic image and the defect mask, performs data analysis, and obtains the defect features of the high-temperature original image.

[0018] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. An artificial intelligence-based high-temperature slab image processing method, characterized in that, The steps include: S1. Obtain the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain a preprocessed image; finally, denoise the preprocessed image by combining guided filtering and wavelet thresholding to obtain the first processed image; S2. Simulate the heat wave distortion effect on the first processed image through elastic transformation to obtain a second processed image, and apply random brightness perturbation to the second processed image to obtain a third processed image; S3. Obtain the normal sheet image, the noise vector of the original high-temperature image, and the sheet defect type, construct a lightweight model to generate a micro-defect mask, and perform random transformation on the third processed image based on the defect mask to obtain a synthetic image; S4. Generate a high-temperature image dataset based on the synthetic image and the defect mask and perform data analysis to obtain the defect features of the high-temperature original image.

2. The method for processing high-temperature slab image based on artificial intelligence according to claim 1, wherein For S1, obtain the original high-temperature image , where H and W are the pixel height and width respectively; the method for dynamically compressing the highlight area of the original high-temperature image includes: Obtain the initial brightness value of each pixel point of the original high-temperature image, and perform dynamic compression on the original high-temperature image based on the initial brightness value. The compression formula is as follows: ； Wherein, is the processed brightness value; I(i, j) is the initial brightness value; D is the maximum value of the standard brightness range; α1 is an adaptive parameter calculated according to the image dynamic range, , μ1 is the mean value of the brightness values of the original high-temperature image, and σ1 is the variance of the square brightness value; Based on the processed brightness value of each pixel , a preprocessed image is generated.

3. The method for processing high-temperature slab image based on artificial intelligence according to claim 2, characterized in that The methods for analyzing and obtaining the separation of the illumination component and the reflection component and enhancing the contrast of the reflection component include: Convolve the preprocessed image with three Gaussian kernels. The convolution formula is as follows: ； In the formula, is the illumination component based on the Gaussian kernel; σ2 is the smoothing parameter used to control the Gaussian kernel, and σ2 ∈ {15, 80, 200}.

4. The method for processing high-temperature slab image based on artificial intelligence according to claim 3, characterized in that, Take the logarithm of the illumination component at each scale and subtract them to obtain the corresponding reflection component at multiple scales , and the calculation formula is as follows: ； The final reflection component R(x) is calculated by fusing and averaging the multi-scale reflection components. The formula is as follows: 。 5. The method for processing high-temperature slab image based on artificial intelligence according to claim 4, characterized in that, Perform enhancement processing on the final reflection component R(x), stretch the histogram of the final reflection component R(x) to [0, D], and apply CLAHE (Contrast Limited Adaptive Histogram Equalization); The methods for combining guided filtering and wavelet thresholding for denoising to obtain the first processed image include: Use the final reflection component R(x) as the guidance map to filter the original high-temperature image, perform 3-layer wavelet decomposition on the filtered image to obtain high-frequency subbands and low-frequency subbands, and apply a global hard threshold to the high-frequency subbands to obtain the first processed image.

6. The method for processing high-temperature slab image based on artificial intelligence according to claim 5, wherein, Regarding S2, the method for simulating the heat wave distortion effect through elastic transformation to obtain the second processed image includes: Create two independent Gaussian filter displacement matrices and , where the size of the Gaussian filter displacement matrix is the same as that of the first processed image. The two Gaussian filter displacement matrices and are calculated based on the following formula: ， ； Wherein, α2 is the displacement amplitude; σ4 is the standard deviation of the Gaussian kernel; is a two-dimensional Gaussian distribution random number; i and j are respectively the abscissa and ordinate of the pixel points in the processed image; Calculate the new coordinates for each pixel point (i, j) of the processed image , and the calculation formula is as follows: ， ； Update the coordinates of each pixel point according to the new coordinates, and resample the processed image after calculating the new coordinates using bilinear interpolation to obtain the second processed image, avoiding the aliasing effect.

7. The method for processing high-temperature slab image based on artificial intelligence according to claim 6, wherein The methods for applying random brightness perturbation to the second processed image to obtain the third processed image include: Extract the brightness channel of the processed image, and perform random scaling on the brightness channel. The random scaling expression is as follows: ； In the formula, is the scaled brightness; V is the original brightness; δ is the random scaling parameter, where δ ∈ (-0.3, 0.3); The methods for constructing a lightweight model to generate a micro-defect mask include: Define the defect type label using one-hot encoding, input the noise vector and the defect type label into the defect generator network to obtain the defect mask M, and perform normalization processing on the defect mask M; Superimpose the defect mask M on the normal sheet image at a random position and a random rotation angle for transformation and scaling to obtain a transformed image. The transformation and scaling are implemented based on the following formula: ； In the formula, is the synthetic image; is the normal sheet image; is the defect image; the symbol "⊙" represents element-wise multiplication.

8. The method for processing high-temperature slab image based on artificial intelligence according to claim 7, characterized in that The methods for learning the defect features of the high-temperature original image include: Load a batch of synthetic images and defect masks M from the high-temperature image dataset, and input the preprocessed MobileNetV3-small model with the synthetic images of the same batch to obtain the original output results. Normalize the original output results to the probability space through the Sigmoid activation function to obtain the defect probabilities. Calculate the classification loss for each pixel point (i, j), and the calculation formula is as follows: ； In the formula, z(i, j, b) is the true mask label, z(i, j, b) ∈ {0, 1}; P(i, j, b) is the pixel defect probability; b is the synthetic image number of the same batch, b is a positive integer, b ∈ [1, B], B is the β batch size; is the positive sample weight; F() is the loss function.

9. The method for processing high-temperature slab image based on artificial intelligence according to claim 8, wherein, The expression of the loss function is as follows: ; where a, b, and c are all variables; Sum and normalize all pixels through the following formula: ; wherein, is the total loss; Take the total loss as the defect feature of the original high-temperature image.

10. An artificial intelligence-based high-temperature slab image processing system, applied to the artificial intelligence-based high-temperature slab image processing method according to any one of the above claims 1 to 9, characterized in that, Specifically include: The image acquisition and first processing module is used to acquire the original high-temperature image of the slab, dynamically compress the highlight area of the original high-temperature image, then separate the illumination component and the reflection component, enhance the contrast of the reflection component to obtain the preprocessed image; finally, denoise the preprocessed image by combining guided filtering and wavelet thresholding to obtain the first processed image; The second processing module is used to simulate the heat wave distortion effect by performing elastic transformation on the first processed image to obtain the second processed image, and apply random brightness perturbation to the second processed image to obtain the third processed image; The third processing module is used to obtain the normal sheet image, the noise vector of the original high-temperature image, and the sheet defect type, construct a lightweight model to generate a micro defect mask, and perform random transformation on the third processed image based on the defect mask to obtain the synthetic image; The image analysis and fourth processing module is used to generate a high-temperature image dataset and perform data analysis based on the synthetic image and the defect mask to obtain the defect characteristics of the high-temperature original image.

Images