A super-resolution reconstruction method for industrial CT images based on deep learning

Abstract

The present invention discloses a super-resolution reconstruction method for industrial CT images based on deep learning, and its application relates to the field of industrial CT image technology, in order to solve the problem of low image resolution caused by equipment limitations. The steps include data acquisition and preprocessing, data enhancement, edge enhancement module construction, convolution downsampling module construction, RRDB residual module construction, upsampling and reconstruction, model training and optimization, and splicing trace processing. By using the present invention, super-resolution reconstruction of multi-material CT images can be achieved, and the resolution and clarity of the images can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial CT images, and in particular to a super-resolution reconstruction method of industrial CT images based on deep learning. Background Art

[0002] Industrial CT technology has been widely used in many industrial fields due to its advantages of non-destructive testing, including but not limited to automobiles, materials, railways, aerospace, etc. In particular, in the fields of aerospace, automobile manufacturing and electronic testing, industrial CT has become an indispensable means of testing. With the rapid development of the manufacturing industry and technological upgrading, the demand for high-precision testing equipment continues to increase, and the market size of industrial CT testing machines continues to expand.

[0003] If the parameters of the X-ray source, such as energy, current, scanning time, etc., are not set properly, the image quality may be degraded. If the detector fails or its performance is unstable, or if the object being measured moves or vibrates, the image may be blurred or have ghosting or other distortions.

[0004] When the object to be measured contains metal parts, due to the strong absorption and scattering of X-rays by metal, radial artifacts, band artifacts or striped artifacts with strong contrast between light and dark may appear in the image. When the amount of projection data is insufficient or the projection angle interval is too wide, undersampling artifacts may occur. This artifact usually manifests itself as the difficulty of reconstructing the image to accurately capture the information at the edge or details of the material. This artifact not only affects the clarity of the image, but may also cover up the real structure inside the object.

[0005] Distortion and artifacts are common problems in industrial CT inspection, and they may appear in a variety of scenarios. How to optimize scanning parameters and use advanced correction algorithms to reduce distortion and artifacts is what needs to be overcome at this stage. Summary of the invention

[0006] The present application provides a super-resolution reconstruction method of industrial CT images based on deep learning, which is used to solve the problems of distortion and artifacts commonly found in industrial CT detection in the prior art, and the occurrence of edge weakening of slices after reconstruction due to the reconstruction algorithm.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A super-resolution reconstruction method for industrial CT images based on deep learning, comprising: a collected low-resolution image, characterized in that the reconstruction method comprises the following steps:

[0009] Step S100: data collection and preprocessing to form low-resolution images, collect CT scan images of various materials (such as metals, ceramics, plastics, rocks, etc.), use analytical methods to reconstruct 2D slices as low-resolution images, and generate high-resolution images as training labels through methods such as bicubic interpolation, traditional image processing or pre-trained models;

[0010] Step S200: performing data enhancement on the low-resolution image;

[0011] Step S300: constructing an edge enhancement module, using the Sobel operator to extract edge features in the x and y directions, and merging the two-dimensional convolution layer and the Sobel convolution block after outputting them in parallel, so as to obtain the edge features of the image more clearly;

[0012] Step S400: construct a convolution downsampling module, use a convolution method with a step size of 2 to achieve downsampling, and use a two-dimensional convolution layer and a Leaky ReLU activation function to construct a convolution layer;

[0013] Step S500: constructing an RRDB residual module, introducing the RRDB residual module in the ESRGAN model, and forming a deep network through 23 RRDB residual networks to fully extract image detail information;

[0014] Step S600: performing upsampling and reconstruction, using a nearest neighbor interpolation convolution module to upsample the feature map to achieve high-resolution image reconstruction, with the number of convolution layers being more than 2;

[0015] Step S700: Enter the super-resolution model training phase, use the Adam optimizer and L1 loss function to train the model, optimize the inference process through cropping and splicing technology, and reduce the memory usage;

[0016] Step S800: reducing the splicing marks and then eliminating the splicing marks.

[0017] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S100, data is collected from a CT scan image of the collected object, and the material of the collected object includes: metal, ceramic, plastic or rock.

[0018] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S100, the CT scan image of the collected object is reconstructed into a two-dimensional image using an analytical reconstruction algorithm;

[0019] The analytical reconstruction algorithm uses a filtered back-projection algorithm.

[0020] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S100, a bicubic interpolation method is used to generate a high-resolution image as a training label.

[0021] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S100, a high-resolution image is generated as a training label by using a traditional image processing method.

[0022] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S100, a pre-trained model is used to generate a high-resolution image as a training label.

[0023] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S200, the first training uses different scale scaling, blurring, flipping, rotation and CutBlur methods to perform data enhancement to improve the generalization ability of the model;

[0024] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S300, the convolution factors of the Sobel operator in the x horizontal direction and the y vertical direction are:

[0025] x-direction:

[0026] -1 0 -1

[0027] -2 0 +2

[0028] -1 0 +1

[0029] y direction:

[0030] +1 +2 +1 0 0 0

[0032] -1 -2 -1

[0033] The convolution factors of the two directions of the CT scan image of the collected object are calculated by plane convolution, and the calculation formula is as follows:

[0034] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S300, the grayscale particle size of each pixel of the CT scan image of the collected object is calculated as follows: , using an approximation replace.

[0035] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S400, the calculation method of downsampling by convolution with a step size of 2 is: , where Out is the output size, In is the input size, pad is the number of border padding of the input matrix, str is the step size, and k is the size of the convolution filter;

[0036] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S400, the convolution layer adopts a two-dimensional convolution in the convolution mode with a step size of 2, and its calculation formula is as follows: , F(x) is the feature map obtained after convolution, and g(x) is the convolution filter.

[0037] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S400, the activation function adopts Leaky ReLU, which improves the "neuron necrosis" in the ReLU activation function and balances the zero gradient problem by changing the linear component in the negative value area before the slope. The expression is: In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step 600, the feature map of each convolutional layer is: F ( x )=LReLU(conv2d(nearest( F ( x ),scale))), scale=2, and each layer mainly enlarges the input feature map by twice.

[0038] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S700, the number of iterations in the training phase is 1000 times, Adam is used as the optimizer to update the model parameters, and the learning rate is set to , the input size of the high-resolution image of this model is 256, and it is randomly cut on the input image, and the batch size is set to 8;

[0039] Using L1 loss function: , where x is the result predicted by the model and y is the true label;

[0040] The peak signal-to-noise ratio (PSNR) is selected as the evaluation index of the model. PSNR mainly calculates the error between the output of the model and the corresponding pixel of the standard high-resolution image. The larger the difference, the closer the generated image is to the label image. The calculation formula is: Where MSE is the root mean square error.

[0041] In the above-mentioned super-resolution reconstruction method of industrial CT images based on deep learning, in step S800, there is an overlapping area between the spliced ​​images, and the overlapping area is set to d, and the splicing expression is:

[0042] According to the above embodiment, a super-resolution reconstruction method of industrial CT images based on deep learning has the following effects:

[0043] 1. Through deep learning technology, super-resolution reconstruction of multi-material CT images is achieved, which improves the resolution and clarity of the images and avoids the problem of low image resolution due to equipment limitations;

[0044] 2. Through data enhancement and edge enhancement modules, the generalization ability of the model and the ability to extract image edge features are improved;

[0045] 3. The RRDB residual module is introduced to fully extract image detail information, further improve the quality of reconstructed images, and reduce noise and artifacts in reconstructed CT images;

[0046] 4. The reasoning process is optimized through cropping and splicing technology, which reduces the memory usage and improves the practicality of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of a super-resolution reconstruction method for industrial CT images based on deep learning;

[0048] Figure 2 A schematic diagram of the RRDB residual module of a super-resolution reconstruction method for industrial CT images based on deep learning, which is derived from the ESRGAN super-resolution model;

[0049] Figure 3 This is a connection diagram of the front and back convolutional layers of a super-resolution reconstruction method for industrial CT images based on deep learning:

[0050] Figure 4 Schematic diagram of the F1 image and F2 image expression of a super-resolution reconstruction method for industrial CT images based on deep learning. DETAILED DESCRIPTION

[0051] In order to make the technical means, creative features, objectives and effects of the invention easy to understand, the present invention is further explained below with reference to specific illustrations.

[0052] like Figure 1 As shown, the present invention provides a super-resolution reconstruction method of industrial CT images based on deep learning, comprising: a collected low-resolution image, characterized in that the reconstruction method comprises the following steps:

[0053] Step S100: data collection and preprocessing to form low-resolution images, collect CT scan images of various materials (such as metals, ceramics, plastics, rocks, etc.), use analytical methods to reconstruct 2D slices as low-resolution images, and generate high-resolution images as training labels through methods such as bicubic interpolation, traditional image processing or pre-trained models;

[0054] In step S100, the CT scan image of the collected object is reconstructed into a two-dimensional image using an analytical reconstruction algorithm; the analytical reconstruction algorithm uses a filtered back-projection algorithm. The filtered back-projection algorithm (FBP) is a reconstruction algorithm widely used in imaging technologies such as computed tomography (CT). The algorithm is based on filtering and back-projection operations of projection data, and can restore high-quality two-dimensional or three-dimensional images from projection data. In the embodiment of the present invention, only two-dimensional images are selected for restoration.

[0055] The following three methods can be used for the selection of training labels, and the preferred method is bicubic interpolation.

[0056] Bicubic interpolation is used to generate high-resolution images as training labels, which is specifically embodied in the following two steps:

[0057] 1. Read low-resolution images;

[0058] image = cv2.imread('low_res_image.jpg');

[0059] 2. Use bicubic interpolation to enlarge the image (2x or 4x);

[0060] scale_factor = 2 or 4;

[0061] resized_image = cv2.resize(image, (0, 0), fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_CUBIC).

[0062] Bicubic interpolation is used here to produce smoother image edges; more information about adjacent pixels is considered during the interpolation process, so the result is usually more accurate than bilinear interpolation;

[0063] Step S200: Perform data enhancement on the low-resolution image and perform the first training; the first training uses different scale scaling, blurring, flipping, rotation and CutBlur methods to perform data enhancement to improve the generalization ability of the model.

[0064] In a specific implementation, a local area of ​​the model input image is a high-resolution image, and the rest is a low-resolution image. Alternatively, a local area is a low-resolution image, and the rest is a high-resolution image. Further cutBlur is an enhancement method here, and the formulas used are as follows:

[0065]

[0066] in, Φ is the area in the image that needs to be cutBlured, h(x,y) is the pixel value at x,y of the high-resolution image, and l(x,y) is the pixel value at x,y of the low-resolution image.

[0067] Going a step further, CutBlur is used to cut low-resolution (LR) patches and paste them into high-resolution (HR) image regions: or vice versa, HR patches are pasted into LR image regions. Doing so allows the model to learn not only how to restore images, but also where to restore them. Sharp transitions or mixed image content in image patches degrade image restoration performance. CutBlur reduces boundary effects by minimizing this mismatch, thereby providing a regularization effect. Since CutBlur has a regularization effect, it not only alleviates the problem of over-sharpening, but also enhances the restoration performance of other LR regions. This is how blur removal is achieved.

[0068] Step S300: construct an edge enhancement module, use the Sobel operator to extract edge features in the x and y directions, and output the two-dimensional convolution layer and the Sobel convolution block in parallel and then merge them to obtain the edge features of the image more clearly. In step S300, the convolution factors of the Sobel operator in the x horizontal direction and the y vertical direction are:

[0069] x-direction:

[0070] -1 0 -1

[0071] -2 0 +2

[0072] -1 0 +1

[0073] y direction:

[0074] +1 +2 +1 0 0 0

[0076] -1 -2 -1

[0077] The grayscale and denoised image is convolved with Gx and Gy respectively to obtain two gradient images (Gx and Gy).

[0078] The convolution factors of the two directions of the CT scan image of the collected object are calculated by plane convolution, and the calculation formula is as follows:

[0079]

[0080] In step S300, the grayscale particle size of each pixel of the CT scan image of the collected object is calculated as follows: .

[0081] The gradient magnitude image G is obtained by calculating the square root of the sum of the squares of Gx and Gy, and using the approximation replace.

[0082] Map the calculated gradient magnitude to a suitable display range (such as 0-255) to better display the edge of the image. In order to highlight the edge, a threshold can be set to display only those pixels whose gradient magnitude exceeds the threshold. Display or save the processed edge-enhanced image.

[0083] In practical applications, it may be necessary to adjust the convolution kernel size, threshold and other parameters of the Sobel operator according to specific needs. For images with more noise, denoising is particularly important, otherwise it may affect the effect of edge detection. The Sobel operator is sensitive to the direction and intensity of the edge, so in some cases it may be necessary to combine other edge detection algorithms for comprehensive analysis. Through the above steps, an edge enhancement module based on the Sobel operator can be constructed, which can effectively highlight the edge information in the image and provide strong support for subsequent image processing tasks.

[0084] Step S400: construct a convolution downsampling module, use a convolution method with a step size of 2 to achieve downsampling, and use a two-dimensional convolution layer and a Leaky ReLU activation function to construct a convolution layer.

[0085] In step S400, the calculation method for downsampling using a convolution method with a stride of 2 is: , where Out is the output size, In is the input size, pad is the number of border padding of the input matrix, str is the step size, and k is the size of the convolution filter;

[0086] out: The size of the output feature map: usually height or width, because height and width are calculated in the same way.

[0087] in: The size of the input feature map: height or width.

[0088] pad: The size of the padding. Padding refers to adding extra zero-valued pixels around the borders of the input feature map to control the size of the output feature map.

[0089] str: The size of the stride. The stride refers to the step length when the convolution kernel moves on the input feature map.

[0090] k: The size of the convolution filter (kernel): usually the height or width of the convolution kernel, assuming the convolution kernel is a square.

[0091] The meaning of the formula (in + 2*pad - k) / str + 1 is:

[0092] in + 2*pad: First, the size of the input feature map is added with twice the size of the padding, because the padding is added on both sides of the input feature map.

[0093] in + 2*pad - k: Then, the kernel size is subtracted from the above result. This step takes into account the portion of the input feature map that the kernel covers, but due to the size of the kernel, not all of the padding will be "seen".

[0094] (in + 2*pad - k) / str: Next, divide the above result by the stride. This step takes into account the stride length of the convolution kernel when moving on the input feature map, thereby determining how many such "blocks" there will be on the output feature map.

[0095] (in + 2*pad - k) / str + 1: Finally, 1 is added because even if the last "block" may not completely cover the rest of the input feature map (depending on whether in + 2*pad - k is divisible by str), the output feature map will still contain this part. What needs to be explained here is that if a padding method other than 'VALID' is used, it can be 'SAME'. But in most convolution operations based on 'VALID' padding, this step will not actually increase the output size, but the result of rounding directly.

[0096] It should be noted that this formula is usually used to calculate the one-dimensional case (i.e. only considering the height or width), and assumes that the input feature map, convolution kernel, and output feature map are all square (i.e. the height and width are equal). For non-square cases, the output size of the height and width needs to be calculated separately. In addition, this formula may need to be adjusted according to the specific convolution operation (such as whether padding is used, the specific value of the stride, etc.).

[0097] Here, in step S400, the convolution layer adopts a two-dimensional convolution in a convolution mode with a step size of 2, and its calculation formula is as follows: , F(x) is the feature map obtained after convolution, and g(x) is the convolution filter.

[0098] In step S400, the activation function uses Leaky ReLU, which improves the "neuron necrosis" in the ReLU activation function and balances the zero gradient problem by changing the linear component in the negative value area before the slope. The expression is:

[0099] Here, when the input of the ReLU function is less than 0, the output is 0, which may cause the neuron to die during the training process (i.e. no longer respond to any data). Leaky ReLU avoids this problem by allowing negative input values ​​to have a small non-zero gradient (i.e. αx), so that the neuron can retain a certain activation value during the training process, thereby alleviating the problem of neuron death. Since Leaky ReLU also has output in the negative range, this allows the model to capture more feature information, thereby increasing the expressiveness of the model. When the gradient of Leaky ReLU is not 0 in the negative range, this helps to speed up the training process because the gradient update will not stop because the input is negative. By allowing negative input values ​​to pass, Leaky ReLU can reduce the sensitivity of the model to the input data, thereby enhancing the robustness of the model.

[0100] like Figure 2-3 As shown, step S500: constructing the RRDB residual module, introducing the RRDB residual module in the ESRGAN model, and forming a deep network through 23 RRDB residual networks to fully extract image detail information. Each residual dense block contains multiple convolutional layers, which are directly added through short connections, so that the network can directly learn the residual information of the image. This design helps to avoid the gradient vanishing problem and improve training efficiency.

[0101] In the figure, β=0.2, and the connection between the front and rear convolutional layers is expressed as Figure 3 shown.

[0102] Step S600: performing upsampling and reconstruction, using a nearest neighbor interpolation convolution module to upsample the feature map to achieve high-resolution image reconstruction, with the number of convolution layers being more than 2;

[0103] The feature map of each convolutional layer is: F ( x )=LReLU(conv2d∗nearest( F ( x ),scale)) , scale=2, and each layer mainly enlarges the input feature map by twice.

[0104] Where: F(x): This usually represents a function that takes an input x and returns an output. In this context, F(x) may represent the output of a neural network layer or the entire network, where x is the input data.

[0105] nearest(F(x),scale): This looks like an attempt to upsample (or scale) F(x), where nearest represents the nearest neighbor interpolation method and scale is the magnification factor. However, typically nearest neighbor interpolation is not applied directly as a function to the output of another function F(x); instead, it is applied to the feature map or image data itself. Also note that the symbol * here is usually used to indicate a convolution operation, not a sign of multiplication or some combination operation.

[0106] conv2d: This represents the 2D convolution operation, which is one of the basic building blocks in convolutional neural networks. 2D convolution is usually applied to feature maps to extract local features.

[0107] LReLU: (Leaky Rectified Linear Unit), which is a nonlinear activation function used to increase the nonlinearity of the network.

[0108] Upsampling: First, we use nearest neighbor interpolation to scale the feature map x to its original size. This step is usually not expressed as a function of F(x), but is applied directly to x.

[0109] 2D Convolution: We then apply a 2D convolution operation to the upsampled feature map to extract features. Upsampling refers to increasing the sampling rate or resolution of a signal or image, making it larger or more detailed. In image processing, upsampling is often used to enlarge an image, increase its size or resolution. Its benefit is that it is achieved through

[0110] Interpolation methods: including nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation. These methods increase the resolution of the image by calculating the values ​​of new pixels to fill in the enlarged image. Interpolation methods estimate the values ​​of new pixels based on the values ​​of surrounding known pixels, so they can retain certain image details.

[0111] Transposed convolution: In neural networks, transposed convolution is a commonly used upsampling method. It upscales low-resolution feature maps to high resolution by reversing the process of convolution operation. Transposed convolution not only increases the size of the image, but also restores the detailed information of the image through the learning ability of the convolution kernel.

[0112] LReLU Activation: Finally, we apply the LReLU activation function to increase the non-linearity of the network.

[0113] It should be explained that in the downsampling in step 400 and the upsampling in step 600, downsampling reduces the computational complexity and extracts high-level features by reducing the resolution of the image, while upsampling restores detail information and improves image clarity by increasing the resolution of the image.

[0114] Step S700: Enter the super-resolution model training phase, use the Adam optimizer and L1 loss function to train the model, optimize the inference process through cropping and splicing technology, and reduce the memory usage.

[0115] The number of iterations in this training phase is 1000, and Adam is used as the optimizer to update the model parameters. The learning rate is set , the input size of the high-resolution image of this model is 256, and it is randomly cut on the input image, and the batch size is set to 8;

[0116] Using L1 loss function: , where x is the result predicted by the model and y is the true label;

[0117] The peak signal-to-noise ratio (PSNR) is selected as the evaluation index of the model. PSNR mainly calculates the error between the output of the model and the corresponding pixel of the standard high-resolution image. The larger the difference, the closer the generated image is to the label image. The calculation formula is:

[0118] Where MSE is the root mean square error.

[0119] Step S800: reducing the splicing traces and then eliminating the splicing traces. There is an overlapping area between the spliced ​​images, and the overlapping area is set to d. The splicing expression is:

[0120] When using the optimal model to deploy inference, due to the large size of CT reconstructed images, directly inputting them into the model for inference will take up a lot of video memory and consume a lot of resources. Scaling the image and inputting it into the model will lose the meaning of super-resolution, and some information will be lost after scaling. In this implementation method, the cropping and splicing methods are used. Using conventional cropping methods, due to the characteristics of convolution, there will be certain errors after using the super-resolution model for inference at the edge, and there will be cropping marks after splicing. This is because the data distribution in the edge area is different. The model uses the method of filling with 0, and there are certain errors in the pixel values ​​of different local edge areas. In this system, a cropping method with repeated edges is used, such as Figure 4As shown, d is the width of the repeated area. The appropriate cropping size can save video memory and save inference time while ensuring accuracy. When stitching, we take the square F1 as the target image pixel value for the first image, and the square F2 as the target image pixel value for the second image. The original side length is w, and d contains the pixels of F1 and F2. The same method is used in the width and height directions. Combined with the above expression, this method can effectively reduce the stitching traces.

[0121] In summary, the present invention provides a super-resolution reconstruction method for industrial CT images based on deep learning. Through deep learning technology, super-resolution reconstruction of multi-material CT images is realized, the resolution and clarity of the image are improved, and the problem of low image resolution due to equipment limitations is avoided. Through data enhancement and edge enhancement modules, the generalization ability of the model and the ability to extract image edge features are improved. The RRDB residual module is introduced to fully extract image detail information, further improve the quality of the reconstructed image, reduce the noise and artifacts of the reconstructed CT image, optimize the reasoning process through cropping and splicing technology, reduce video memory occupancy, and improve the practicability of the model.

[0122] The above describes the specific embodiments of the invention. It should be understood that the invention is not limited to the above specific embodiments, and the devices and structures not described in detail should be understood to be implemented in a common manner in the art; those skilled in the art can make various modifications or modifications within the scope of the claims, and make some simple deductions, deformations or substitutions, which does not affect the essence of the invention.

Claims

1. A super-resolution reconstruction method for industrial CT images based on deep learning, comprising: Collecting CT scan images of various materials, characterized in that the reconstruction method includes the following steps: Step S100: collecting and preprocessing data to form a low-resolution image; Step S200: performing data enhancement on the low-resolution image; Step S300: constructing an edge enhancement module; Step S400: construct a convolution downsampling module, use a convolution method with a step size of 2 to achieve downsampling, and use a two-dimensional convolution layer and a Leaky ReLU activation function to construct a convolution layer; Step S500: constructing an RRDB residual module, introducing the RRDB residual module in the ESRGAN model, and forming a deep network through 23 RRDB residual networks to fully extract image detail information as a feature map; Step S600: performing upsampling and reconstruction, using a nearest neighbor interpolation convolution module to upsample the feature map to achieve high-resolution image reconstruction, with the number of convolution layers being more than 2; Step S700: Enter the super-resolution model training phase, use the Adam optimizer and L1 loss function to train the model, optimize the inference process through cropping and splicing technology, and reduce the memory usage; Step S800: reducing splicing marks, and then eliminating splicing marks; In step S300, the convolution factors of the Sobel operator in the x horizontal direction and the y vertical direction are: x-direction: ； y direction: ； Planar convolution is performed on the convolution factors of the two directions of the CT scan image of the collected object. A is the CT scan image of the collected object. The calculation formula is as follows: ； The grayscale particle size of each pixel of the CT scan image of the collected object is calculated as follows: , using an approximation replace; In step S400, the calculation method of downsampling by convolution with a step size of 2 is: , where Out is the output size, In is the input size, pad is the number of border padding of the input matrix, str is the step size, and k is the size of the convolution filter; In step S400, the convolution layer uses a two-dimensional convolution in the convolution mode with a step size of 2, and its calculation formula is as follows: , F(x) is the feature map obtained after convolution, and g(x) is the convolution filter; In step S400, the activation function uses Leaky ReLU, which improves the "neuron necrosis" in the ReLU activation function and balances the zero gradient problem by changing the linear component in the negative value area before the slope. The expression is: ； In step S600, the feature map of each convolutional layer is: F ( x )=LReLU(conv2d(nearest( F ( x ),scale))), scale=2, and each layer magnifies the input feature map by twice; In step S700, the number of iterations in this training phase is 1000, and Adam is used as the optimizer to update the model parameters, and the learning rate is set to , the input size of the low-resolution image of this model is 256, and it is randomly cut on the input image, and the batch size is set to 8; Using L1 loss function: , where x is the result predicted by the model and y is the true label; The peak signal-to-noise ratio (PSNR) is selected as the evaluation index of the model. PSNR is the error between the output of the calculation model and the corresponding pixel of the standard high-resolution image. The larger the difference of the peak signal-to-noise ratio (PSNR), the closer the generated image is to the label image. The calculation formula is: ； Where MSE is the root mean square error, There is an overlapping area between the stitched images, which is set to d. The width of the stitched image is W. The stitching expression is: 。 2. The method for super-resolution reconstruction of industrial CT images based on deep learning according to claim 1, characterized in that: In step S100, data is collected from a CT scan image of an object to be collected, and the material of the object to be collected includes: metal, ceramic, plastic or rock; The CT scan image of the collected object is subjected to an analytical reconstruction algorithm to generate a two-dimensional image reconstruction; the analytical reconstruction algorithm adopts a filtered back-projection algorithm.

3. The super-resolution reconstruction method of industrial CT images based on deep learning according to claim 2, characterized in that: In step S100, a high-resolution image is generated as a training label by using bicubic interpolation.

4. The method for super-resolution reconstruction of industrial CT images based on deep learning according to claim 2, characterized in that: In step S100, a high-resolution image is generated as a training label using a traditional image processing method.

5. The method for super-resolution reconstruction of industrial CT images based on deep learning according to claim 2, characterized in that: In step S100, a pre-trained model is used to generate high-resolution images as training labels.

6. The method for super-resolution reconstruction of industrial CT images based on deep learning according to any one of claims 3, 4 or 5, characterized in that: In step S200, data enhancement is performed by using different scale scaling, blurring, flipping, rotating and CutBlur methods to improve the generalization ability of the model.

Images