CT single frame multi-frame super-resolution fusion image enhancement method and 3D online detection machine

Abstract

The application discloses a CT single frame multi-frame super-resolution fusion image enhancement method, which comprises the following steps: S1, establishing a projection number-image effect model; S2, establishing a noise joint projection number-image effect model; S3, based on the established projection number-image effect model and noise joint projection number-image effect model, degrading a CT image set to obtain a degraded CT image set; S4, constructing a CT image inversion model based on a convolutional neural network, using the CT image before degradation, the degraded CT image and the adjacent slice image after degradation as the training data of the network, training the CT image inversion model, and then based on the trained CT image inversion model, performing image enhancement on the CT image to be processed. The application can effectively reduce image noise, enhance details in the image and convert industrial CT slice images into high-quality clear images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for enhancing CT single-frame and multi-frame super-resolution fusion images and a 3D online inspection machine. Background Technology

[0002] In the field of industrial CT imaging, image quality is not only affected by the multi-scale characteristics of the internal structure of the object being inspected, but also severely interfered with by the physical limitations of the imaging system itself (such as X-ray scattering, electronic noise, and quantum noise). In production line environments where efficiency is paramount, due to UPH efficiency requirements, CT image acquisition speed is limited to only 3 seconds or even less, and the number of projected images acquired is often no more than 64. This results in a massive amount of noise in CT images when dealing with products with poor penetration. This noise not only severely masks the true defect information but is also frequently misjudged as a defect, causing a large number of false positives in the detection system. For products with high density, complex materials, or uneven thickness, X-ray penetration ability is significantly reduced, and the noise problem becomes even more prominent. Tiny defects such as bubbles and holes in the image are almost indistinguishable from background noise, greatly interfering with the accuracy of the detection algorithm.

[0003] like Figure 4 The image shown is 512 projected images of a BGA product. Figure 5 With 32 projected images, the impact of the number of projected images on image quality is clearly visible: Figure 4 It clearly has a lower signal-to-noise ratio, while in contrast... Figure 5 It can be seen that the number of projected images affects the image noise and edge sharpness.

[0004] Currently, industrial CT images typically require preprocessing and enhancement to improve the effectiveness of subsequent human observation and automated inspection. Common image processing workflows include noise reduction and detail enhancement. For noise reduction, traditional methods such as Gaussian filtering, median filtering, and mean filtering are widely used, while sharpening commonly employs methods such as desharpening masks and various convolution kernels. However, these traditional algorithms often require manual parameter adjustment based on the actual image quality and are more applicable to 2D images, limiting their adaptability and automation.

[0005] Among them, patent document CN108242066B provides a spatial resolution enhancement device and method for CT images, as well as a CT imaging system. The method includes: acquiring the original CT projection curve; performing deconvolution on the projection data on the original CT projection curve in the X-ray tube sampling direction or the texture direction of the original CT projection curve; and reconstructing the image based on the projection data after deconvolution. However, the convolution kernel of this method is a pre-set empirical value, which cannot be adapted to different application scenarios. Patent document CN119067853B invented a super-resolution reconstruction method for industrial CT images based on deep learning, applicable to the field of industrial CT image technology, to solve the problem of low image resolution due to equipment limitations. However, this method uses bicubic interpolation to generate high-resolution images as training labels. This interpolation method introduces noise into the generated high-resolution images and limits the capabilities of the neural network. Patent document CN202411478424.9 provides an enhanced SwinTransformer super-resolution reconstruction method for industrial CT images, belonging to the fields of computer image processing and industrial non-destructive testing. However, this method only uses a single CT image as network input and does not consider the contextual information between slices.

[0006] In summary, the methods mentioned in the above patents all have certain shortcomings, so it is necessary to improve them. Summary of the Invention

[0007] The purpose of this invention is to provide a method for enhancing CT single-frame and multi-frame super-resolution fusion images and a 3D online inspection machine, so as to effectively reduce image noise, enhance the details in the image, and convert industrial CT slice images into high-quality and clear images.

[0008] To achieve the above objectives, the following technical solution is adopted:

[0009] A method for enhancing single-frame and multi-frame super-resolution fusion images in CT scans includes the following steps:

[0010] S1: Establish a projection number-image quality model;

[0011] S2: Establish a noise joint projection number-image effect model;

[0012] S3: Based on the constructed projection number-image effect model and noise joint projection number-image effect model, the CT image set is degraded to obtain the degraded CT image set;

[0013] S4: Construct a CT image inversion model based on a convolutional neural network, and use the CT image before degradation, the CT image after degradation, and the adjacent slice image after degradation as training data for the network to train the CT image inversion model. Then, based on the trained CT image inversion model, perform image enhancement on the CT image to be processed.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] S11: Generate a random Gaussian kernel, where the size of the Gaussian kernel is (2k+1)×(2k+1), and the calculation formula for each element of the Gaussian kernel is as follows:

[0016] Where i and j represent the indices of the kernel matrix, and δ is the standard deviation;

[0017] S12: Normalized Gaussian kernel, the specific formula is:

[0018] Where G′(i,j) is the normalized Gaussian kernel;

[0019] S13: Apply Gaussian blur to CT images based on a Gaussian kernel. The specific formula is as follows:

[0020] Where I is the original image, I blur The image is Gaussian blurred, where x and y are the coordinates of the image, and i and j represent the indices of the kernel matrix.

[0021] Furthermore, step S2 specifically includes the following steps:

[0022] S21: Use Poisson noise to simulate quantum noise, where the probability mass function of the Poisson distribution is as follows: Where x is the observed photon count, λ is the expected value of the photon count, and N is the number of selected images, ranging from 32 to 1024.

[0023] S22: Add Poisson noise to the image. Simultaneously, the pixel grayscale values ​​of the image after adding Poisson noise need to be limited. The specific formula is as follows:

[0024] Among them, I d The image after Poisson noise has been added;

[0025] S23: Add salt-and-pepper noise to the image using the following formula:

[0026] Among them, I p For the image with added salt-and-pepper noise, p pepper and p salt This represents the salt-and-pepper noise coefficient.

[0027] Furthermore, step S3 specifically includes the following steps:

[0028] S31: Based on the projection number-image effect model, randomly select the standard deviation σ of the Gaussian function and the kernel radius k of the Gaussian blur to perform blur degradation on the image;

[0029] S32: For the blurred and degraded image I blur We use a noise joint projection number-image effect model to add quantum noise and salt-and-pepper noise to the image.

[0030] Furthermore, in S4, when constructing the CT image inversion model based on a convolutional neural network, it is necessary to improve the input layer of the RRDBNet network of BSRGAN. At the same time, a cross-slice attention module is added before the RRDB backbone network, and the upsampling layer in the original BSRGAN is replaced with a multi-scale feature enhancement module. The improvement of the input layer is to expand the single image input to receive 2N+1 slices, so that the network can obtain the contextual information between slices. The cross-slice attention module is used to adaptively learn the relationship weights between different slices, highlight the important information related to the central slice, and filter out irrelevant information. The multi-scale feature enhancement module is used to extract multi-scale features in parallel through convolutional kernels of different sizes, and then combine the channel attention mechanism to enhance image details without changing the spatial resolution.

[0031] Furthermore, the 2N+1 slices include the central slice and its N adjacent slices above and below it.

[0032] Furthermore, in S4, the loss function of the constructed CT image inversion model includes perceptual loss, adversarial loss, and content loss; wherein, the perceptual loss is calculated by calculating the feature distance between the enhanced slice and the real high-quality slice, and these features are extracted through a pre-trained VGG network; the adversarial loss is measured by training an auxiliary discriminator network to measure the distribution difference between the generated image and the real high-quality image; the content loss is measured by comparing the pixel-level differences between the generated image and the real high-quality image to measure similarity.

[0033] Furthermore, in step S4, the construction of the CT image inversion model training data specifically includes the following steps:

[0034] Acquire high-quality 3D volumetric data from industrial CT images;

[0035] Extract continuous slice sequences from 3D volume data, where each training sample contains 2N+1 continuous slices;

[0036] CT imaging degradation of the central slices produces low-quality central slices. The specific formula is as follows:

[0037] i l ow=D(i high) = Blur(I h igh,σ b lur,r b lur)+SaltPepperNoise(I h igh,p)+Noise(λ p oisson), where I h igh is the original high-quality slice, I l ow is the degenerate slice, σ b lur is a randomly assigned Gaussian kernel standard deviation, ranging from [0.2, 3], r b lur is a randomly assigned Gaussian kernel radius, ranging from [5, 25], p is the probability parameter of salt-and-pepper noise, ranging from [0.01, 0.1], and λ p oisson is a randomly assigned Poisson noise parameter, with a value range of [0.1, 3].

[0038] Furthermore, in S4, when training the CT image inversion model, the degraded center slice and its adjacent upper and lower slices are used as the input to the model, and the original high-quality center slice is used as the target output of the model. At the same time, the Adam optimizer is also used for optimization, wherein the initial learning rate is set to 0.0002, β1 = 0.9, β2 = 0.999, and the epoch is set to 100000. In addition, in each training iteration, a small batch of slice sequences is first randomly sampled from the dataset, then the loss is calculated through forward propagation, and the model parameters are updated through backpropagation.

[0039] A 3D online inspection machine is also provided, including a product conveying device, an X-ray emitting device, and an X-ray receiving device, which uses the above-mentioned image enhancement method to enhance the acquired product images.

[0040] By adopting the above solution, the beneficial effects of the present invention are:

[0041] This invention, based on the principles of CT reconstruction, constructs a CT slice image degradation model and establishes a CT reconstruction inversion model based on convolutional neural networks to enhance slice images. Compared with traditional image enhancement algorithms, it has better robustness to CT slice images, and the enhancement effect is significantly better than traditional image enhancement algorithms and 2D super-resolution algorithms. At the same time, this invention has low limitation on the number of input slice images, and can input the actual number of slice images (1-21 images) according to actual needs. It also has good results for single input images and has good versatility. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a schematic diagram of the modified RRDBNet network structure of the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of the 3D online inspection machine of the present invention;

[0045] Figure 4 Reconstruct slice images from 512 projected images of the BGA chip;

[0046] Figure 5 Reconstruct slice images from 32 projected images of a BGA chip;

[0047] Figure 6 To use the method of the present invention to Figure 5 Enhanced result image;

[0048] Figure 7 This is a structural diagram of the solder layer of one of the IGBT circuit boards;

[0049] Figure 8 This is a structural diagram of the solder layer on another IGBT circuit board;

[0050] Figure 9 To use the method of the present invention to Figure 7 Enhanced result image;

[0051] Figure 10 To use the method of the present invention to Figure 8 Enhanced result image;

[0052] The following are explanations of the labels in the attached diagram:

[0053] 1. Product conveying device; 2. X-ray emitting device; 3. X-ray receiving device. Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0055] Reference Figures 1 to 10 As shown, this invention provides a method for enhancing CT single-frame and multi-frame super-resolution fusion images, comprising the following steps:

[0056] S1: Establish a projection number-image effect model.

[0057] The number of projection frames in a CT image affects the sharpness of the sliced ​​image. A larger number of projection frames results in less noise and sharper edges, while a smaller number of projection frames leads to increased image noise and blurred edges, thus reducing image sharpness. Typically, edge blurring can be described using Gaussian blur, and the standard deviation of Gaussian blur is related to the number of projection frames, as follows:

[0058] σ g=γ / σ s In the formula, σ s σ is the number of projections. g Let γ be the standard deviation of the Gaussian blur, and γ be the transformation coefficient, which is related to the ratio of the pixel size of the flat panel detector to the distance from the light tube to the object being detected, and the distance from the object being detected to the flat panel detector. Therefore, different Gaussian blur coefficients can be used to apply Gaussian blur to the image to simulate the effect of different projection numbers on image sharpness. The specific steps are as follows:

[0059] S11: Generate a random Gaussian kernel. For an image, the Gaussian kernel is a matrix generated based on a normal distribution. Gaussian kernels are widely used in image processing for smoothing and noise reduction. Assuming the size of the Gaussian kernel is (2k+1)×(2k+1), the calculation formula for each element of the Gaussian kernel is as follows:

[0060] Where i and j represent the indices of the kernel matrix, and δ is the standard deviation;

[0061] S12: Normalized Gaussian kernel. To ensure that the filtered pixel values ​​are within a reasonable range, the Gaussian kernel G(i,j) needs to be normalized so that the sum of all its elements is 1. The specific formula is as follows:

[0062] Where G′(i,j) is the normalized Gaussian kernel;

[0063] S13: Apply Gaussian blur to CT images based on a Gaussian kernel. The specific formula is as follows:

[0064] Where I is the original image, I blur The image is Gaussian blurred, where x and y are the coordinates of the image, and i and j represent the indices of the kernel matrix.

[0065] In summary, the projection number-image quality model can be described by the following formula:

[0066] I blur =f spot (I,δ,k), where δ is the standard deviation and k is the kernel radius. For image I, random values ​​are taken for δ and k to simulate the effect of different projection numbers on image sharpness.

[0067] S2: Establish a noise joint projection number-image effect model.

[0068] In X-ray imaging systems, the main component of random noise in X-ray images is quantum noise, which originates from the discrete nature of X-rays and the photon counting by the flat-panel detector. During X-ray imaging, X-rays are used to irradiate the target object and interact with it. When X-rays interact with atoms inside the object, they are absorbed or scattered, producing one or more photons. These photons are captured by the detector and converted into electrical signals. Because the number of X-rays is discrete and the generation of photons is a random process, the photon count observed in the image is affected by statistical fluctuations. Simultaneously, the number of CT projection images also affects image noise; a larger number of projection images results in less noise and more accurate edges. The specific steps are as follows:

[0069] S21: Use Poisson noise to simulate quantum noise, where the probability mass function of the Poisson distribution is as follows:

[0070] Where x is the observed photon count, λ is the expected value of the photon count, and N is the number of selected images, ranging from 32 to 1024. The larger the λ is, the smaller the image noise.

[0071] S22: Add Poisson noise to the image.

[0072] For scenarios involving adding Poisson distribution noise to image grayscale, by setting an appropriate value for λ, a noisy image I following a Poisson distribution can be generated based on the above formula. n Then it is superimposed on image I, specifically:

[0073] I d =I+I n In the formula, I d The image shows the result after adding Poisson noise. Furthermore, to prevent pixel grayscale values ​​from overflowing in the resulting image, it is necessary to limit the pixel grayscale values ​​of the image after adding Poisson noise. The specific formula is as follows:

[0074] Among them, I d The image after Poisson noise has been added;

[0075] S23: Add salt-and-pepper noise to the image using the following formula:

[0076] Among them, I p For the image with added salt-and-pepper noise, p pepper and p salt This represents the salt-and-pepper noise coefficient.

[0077] In summary, the noise joint projection number-image effect model can be described as follows:

[0078] Ip =f noise (I,λ,p pepper ,p salt (,N), using different parameters λ, p pepper p salt N can simulate different levels of image signal-to-noise ratio.

[0079] S3: Based on the constructed projection number-image effect model and noise joint projection number-image effect model, the CT image set is degraded to obtain the degraded CT image set.

[0080] Based on the two relational models constructed above, the original image is first blurred and degraded using the projection number-image effect model. Then, the noise joint projection number-image effect model is used to add quantum noise and salt-and-pepper noise to the blurred and degraded image. Specifically:

[0081] S31: Based on the projection number-image effect model, a suitable standard deviation σ of the Gaussian function and the kernel radius k of the Gaussian blur are randomly selected to perform blur degradation on the image. The formula is as follows:

[0082] I blur =f spot (I,δ,k), where I is the high-resolution original image, I blur This is a blurred and degraded image.

[0083] S32: For the blurred and degraded image I blur Using the noise joint projection number-image effect model, quantum noise and salt-and-pepper noise are added to the image, with the following formula:

[0084] I p =f noise (I blur ,λ,p pepper, p salt N select N all ), where I p For I blur The result image after adding quantum noise and salt-and-pepper noise.

[0085] S4: Construct a CT image inversion model based on a convolutional neural network, and use the CT image before degradation, the CT image after degradation, and the adjacent slice image after degradation as training data for the network to train the CT image inversion model. Then, based on the trained CT image inversion model, perform image enhancement on the CT image to be processed.

[0086] In S4, when constructing a CT image inversion model based on a convolutional neural network, the input layer of the RRDBNet network of BSRGAN needs to be improved. At the same time, a cross-slice attention module is added before the RRDB backbone network, and the upsampling layer in the original BSRGAN is replaced with a multi-scale feature enhancement module. The improvement of the input layer is to expand the single image input to receive 2N+1 slices, so that the network can obtain contextual information between slices. The cross-slice attention module is used to adaptively learn the relationship weights between different slices, highlight the important information related to the central slice, and filter out irrelevant information. The multi-scale feature enhancement module is used to extract multi-scale features in parallel through convolutional kernels of different sizes, and then combine it with a channel attention mechanism to enhance image details without changing the spatial resolution.

[0087] The 2N+1 slices include the central slice and its N adjacent slices above and below it.

[0088] BSRGAN is a generative adversarial network for image super-resolution, specifically designed to handle complex image degradation. Compared to traditional super-resolution algorithms, BSRGAN introduces a practical image degradation model capable of simulating various complex image degradation processes, including blurring, downsampling, and noise reduction. However, when using BSRGAN to process 3D image slices, the network cannot fully utilize the contextual information between slices, and to meet the super-resolution function, the network includes upsampling layers, making it unsuitable for directly enhancing CT images of the same size. Therefore, the generative network of BSRGAN needs to be modified to be suitable for CT image slice enhancement, specifically:

[0089] 1) Construction of generative models

[0090] Because the BSRGAN model uses upsampling layers (nearest neighbor interpolation and convolution) for image super-resolution tasks, while industrial CT image slice enhancement tasks require maintaining consistent input and output sizes and fully utilizing spatial relationship information between adjacent slices, the generative network needs improvement. Considering the special characteristics of industrial CT images, this invention designs a novel Enhanced3DSliceNet network structure, which can receive multiple adjacent slices as input and output high-quality center slice enhancement results. Specifically, three key improvements are made to the BSRGAN RRDBNet network:

[0091] Input layer improvement: The single image input is expanded to receive 2N+1 slices (the center slice and its N adjacent slices above and below it). In this way, the network can obtain contextual information between slices.

[0092] Cross-slice attention module (CSAM): A cross-slice attention module is added before the RRDB backbone network. This module adaptively learns the weights of the relationships between different slices, highlighting important information related to the central slice and filtering out irrelevant information.

[0093] Multi-scale feature enhancement module: The upsampling layer in the original BSRGAN is replaced by a multi-scale feature enhancement module. This module extracts multi-scale features in parallel through convolutional kernels of different sizes and combines channel attention mechanism to enhance image details without changing spatial resolution.

[0094] 2) Loss Function

[0095] The loss function includes perceptual loss, adversarial loss, and content loss. The perceptual loss is calculated by measuring the feature distance between the enhanced slice and the real high-quality slice, and these features are extracted through a pre-trained VGG network. The adversarial loss measures the distribution difference between the generated image and the real high-quality image by training an auxiliary discriminator network. The content loss measures similarity by comparing the pixel-level differences between the generated image and the real high-quality image. In this invention, L1 loss is selected as the content loss function.

[0096] 3) Training data construction

[0097] The construction of training data for the CT image inversion model includes the following steps:

[0098] Acquire high-quality industrial CT image 3D volumetric data (CTSpine1K);

[0099] Extract continuous slice sequences from 3D volume data, where each training sample contains 2N+1 continuous slices;

[0100] CT imaging degradation of the central slices produces low-quality central slices. The specific formula is as follows:

[0101] I l ow=D(I h igh) = Blur(I h igh,σ b lur,r b lur)+SaltPepperNoise(I h igh,p)+Noise(λ p oisson), where I h igh is the original high-quality slice, I l ow is the degenerate slice, σ b lur is a randomly assigned Gaussian kernel standard deviation, ranging from [0.2, 3], r blur is a randomly assigned Gaussian kernel radius, ranging from [5, 25], p is the probability parameter of salt-and-pepper noise, ranging from [0.01, 0.1], and λ p oisson is a randomly assigned Poisson noise parameter, with a value range of [0.1, 3].

[0102] 4) Model Training

[0103] When training the CT image inversion model, the degraded center slice and its adjacent upper and lower slices are used as the model input, and the original high-quality center slice is used as the model output. The Adam optimizer is also used for optimization, with an initial learning rate of 0.0002, β1 = 0.9, β2 = 0.999, and epochs of 100,000. In each training iteration, a small batch of slice sequences is randomly selected from the dataset, and the loss is calculated through forward propagation, followed by updating the model parameters through backpropagation. The training process consists of two stages: the first stage uses only content loss to train the base network, optimizing the network's ability to restore image content; the second stage introduces perceptual loss and adversarial loss to further improve the visual quality of the images. The training process automatically terminates early when the performance metrics on the validation set show no significant improvement for 10 consecutive epochs.

[0104] The feasibility of this invention has not been verified. Figure 6 To use the above method to Figure 5 The enhanced result of the BGA image shows that... Figure 6 The noise level is significantly reduced, while the BGA edges are more pronounced; Figure 7 , Figure 8 All of these are solder layers on IGBT circuit boards. Figure 9 To use the inversion model for Figure 7 The enhanced result of the IGBT solder pattern shows that noise is greatly suppressed and bubble edges are clearer; Figure 10 To use the inversion model for Figure 8 The enhanced result of the IGBT solder pattern shows a significant reduction in noise and a marked increase in the contrast between bubbles and the background.

[0105] In addition, a 3D online inspection machine is provided, including a product conveying device, an X-ray emitting device, and an X-ray receiving device, which uses the above-mentioned image enhancement method to enhance the acquired product images.

[0106] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing CT single-frame and multi-frame super-resolution fusion images, characterized in that, Includes the following steps: S1: Establish a projection number-image quality model; S2: Establish a noise joint projection number-image effect model; S3: Based on the constructed projection number-image effect model and noise joint projection number-image effect model, the CT image set is degraded to obtain the degraded CT image set; S4: Construct a CT image inversion model based on a convolutional neural network, and use the CT image before degradation, the CT image after degradation, and the adjacent slice image after degradation as training data for the network to train the CT image inversion model. Then, based on the trained CT image inversion model, perform image enhancement on the CT image to be processed. In S4, when constructing a CT image inversion model based on a convolutional neural network, the input layer of the RRDBNet network of BSRGAN needs to be improved. At the same time, a cross-slice attention module is added before the RRDB backbone network, and the upsampling layer in the original BSRGAN is replaced with a multi-scale feature enhancement module. The improvement of the input layer is to expand the single image input to receive 2N+1 slices, so that the network can obtain contextual information between slices. The cross-slice attention module is used to adaptively learn the relationship weights between different slices, highlight the important information related to the central slice, and filter out irrelevant information. The multi-scale feature enhancement module is used to extract multi-scale features in parallel through convolutional kernels of different sizes, and then combine it with a channel attention mechanism to enhance image details without changing the spatial resolution.

2. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 1, characterized in that, S1 specifically includes the following steps: S11: Generate a random Gaussian kernel, where the size of the Gaussian kernel is... For each element of the Gaussian kernel, the calculation formula is as follows: ,in, and Indicates the index of the kernel matrix. Standard deviation; S12: Normalized Gaussian kernel, the specific formula is: ,in, The normalized Gaussian kernel; S13: Apply Gaussian blur to CT images based on a Gaussian kernel. The specific formula is as follows: ,in, This is the original image. The image after Gaussian blurring. and These are the coordinate values ​​of the image. and Indicates the index of the kernel matrix.

3. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 2, characterized in that, S2 specifically includes the following steps: S21: Use Poisson noise to simulate quantum noise, where the probability mass function of the Poisson distribution is as follows: ,in, It is the observed photon count. It is the expected value of the photon count. This represents the number of images selected, ranging from 32 to 1024. S22: Add Poisson noise to the image. Simultaneously, the pixel grayscale values ​​of the image after adding Poisson noise need to be limited. The specific formula is as follows: ,in, The image after Poisson noise has been added; S23: Add salt-and-pepper noise to the image using the following formula: ,in, This is the image after adding salt and pepper noise. and This represents the salt-and-pepper noise coefficient.

4. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 3, characterized in that, S3 specifically includes the following steps: S31: Based on the projection number-image quality model, randomly select the standard deviation of a suitable Gaussian function. kernel radius with Gaussian blur To blur and degrade the image; S32: For images that have been blurred or degraded We use a noise joint projection number-image effect model to add quantum noise and salt-and-pepper noise to the image.

5. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 1, characterized in that, The 2N+1 slices include the central slice and its N adjacent slices above and below it.

6. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 1, characterized in that, In step S4, the loss function of the constructed CT image inversion model includes perceptual loss, adversarial loss, and content loss. The perceptual loss is calculated by measuring the feature distance between the enhanced slice and the real high-quality slice, and these features are extracted by a pre-trained VGG network. The adversarial loss is measured by training an auxiliary discriminator network to measure the distribution difference between the generated image and the real high-quality image. The content loss is measured by comparing the pixel-level differences between the generated image and the real high-quality image to measure similarity.

7. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 1, characterized in that, In step S4, the construction of the CT image inversion model training data specifically includes the following steps: Acquire high-quality 3D volumetric data from industrial CT images; Extract continuous slice sequences from 3D volume data, where each training sample contains 2N+1 continuous slices; CT imaging degradation of the central slices produces low-quality central slices. The specific formula is as follows: ,in, For original high-quality slices, The slide is a degenerated section. The standard deviation of the Gaussian kernel is randomly assigned, with a value range of [0.2, 3]. The radius of the Gaussian kernel is randomly assigned, with values ​​ranging from [5, 25]. is the probability parameter for salt-and-pepper noise, with a value range of [0.01, 0.1]. The Poisson noise parameter is randomly assigned and its value ranges from [0.1, 3].

8. The CT single-frame and multi-frame super-resolution fusion image enhancement method according to claim 1, characterized in that, In step S4, when training the CT image inversion model, the degraded center slice and its adjacent upper and lower slices are used as the model input, and the original high-quality center slice is used as the model output. Simultaneously, the Adam optimizer is used for optimization, with an initial learning rate of 0.0002. =0.9, =0.999, epoch is set to 100000. In addition, in each training iteration, a small batch of slice sequences is randomly drawn from the dataset, the loss is calculated through forward propagation, and the model parameters are updated through backpropagation.

9. A 3D online inspection machine, comprising a product conveying device, an X-ray emitting device, and an X-ray receiving device, characterized in that, The image enhancement method described in any one of claims 1 to 8 is used to enhance the acquired product image.

Images