Welding seam X-ray image enhancement method based on generative adversarial network
By using the weld X-ray image enhancement model based on the generative adversarial network, using the generator and discriminator network to process low-resolution images, the problem of difficulty in obtaining high-resolution weld X-ray images in the prior art is solved, and image quality is achieved and improve detection accuracy.
Patent Information
- Application Number
- CN202411101099.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-12
AI Technical Summary
It is difficult to obtain high-resolution weld X-ray images in the prior art, and traditional image processing methods rely on manual experience and cannot effectively deal with the complex and changeable weld X-ray image quality problems.
Design a weld X-ray image enhancement model based on a generative adversarial network. By constructing a data set containing high-resolution and low-resolution images, the image quality is improved by using generator and discriminator networks. The generator network uses convolution modules and residuals. The difference module and the deconvolution module process low-resolution images, and the discriminator network extracts feature information through the convolution module and feature elimination module, and optimizes image quality with attention mechanism and loss function.
The quality of X-ray images of different welds is improved, high-resolution images are generated, the clarity and resolution of images are improved, manual intervention is reduced, and detection accuracy and reliability are improved.
Smart Images

Figure CN118982474A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a weld X-ray image enhancement method based on a generative adversarial network, and belongs to the technical field of weld X-ray image enhancement. Background Art
[0002] In weld X-ray imaging, factors such as equipment resolution, sensor quality, and exposure time may affect the clarity of the final image. Due to these limitations, the acquired images often fail to meet the resolution requirements required for practical applications. At the same time, the image quality may also be lost during transmission and storage due to factors such as compression and noise interference. These losses will further reduce the resolution and clarity of the image, making it difficult to identify the details in the image. In many practical applications, high-resolution images are essential for accurately identifying and analyzing targets. For example, in weld inspection, high-resolution images can more clearly display defects in the weld, thereby improving the accuracy and reliability of inspection.
[0003] However, obtaining high-resolution weld X-ray images is usually limited by the number and performance of equipment. On the one hand, high-precision X-ray imaging equipment and sensors are expensive and may not be suitable for large-scale deployment or general-purpose commercial applications. On the other hand, even with high-performance equipment, it is often difficult to directly obtain ideal high-resolution images due to constraints such as the acquisition environment, network transmission media and bandwidth. Therefore, under the conditions of equipment performance and quantity limitations, it becomes particularly important to achieve quality enhancement of weld X-ray images through algorithms. Due to the different quality of weld X-ray images, traditional image processing methods usually require manual design and adjustment of model parameters, and rely more on threshold selection and manual experience, which not only increases the workload and time cost, but also may not be able to fully cope with various complex and changeable weld X-ray images. Summary of the invention
[0004] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, and propose a weld X-ray image enhancement method based on a generative adversarial network. By designing a weld X-ray image enhancement model based on a generative adversarial network, the quality of different weld X-ray images is improved, and a high-resolution weld X-ray image is obtained.
[0005] The technical solution of the present invention is:
[0006] A weld X-ray image enhancement method based on a generative adversarial network, comprising:
[0007] Collect high-resolution weld X-ray image samples with a spatial resolution of 1 to 20 microns, pre-process each high-resolution weld X-ray image in the sample to obtain a low-resolution weld X-ray image, where the spatial resolution of the low-resolution weld X-ray image is 200 to 400 microns; construct a data set of high-resolution weld X-ray images and low-resolution weld X-ray images containing corresponding relationships;
[0008] A weld X-ray image enhancement model based on generative adversarial network is constructed, in which the generator network input is a low-resolution weld X-ray image and the output is a generated image; the discriminator network input is a high-resolution weld X-ray image and the generated image, and the output is the discriminant probability of whether the generated image is a high-resolution weld X-ray image;
[0009] The data set is used to train the weld X-ray image enhancement model; the low-resolution weld X-ray image that needs to be enhanced in image quality is input into the trained weld X-ray image enhancement model, and the generated image output by the generator network is the enhanced weld X-ray image.
[0010] Preferably, the generator network includes multiple convolution modules, multiple residual modules, and multiple deconvolution modules; the low-resolution weld X-ray image is processed by each convolution module and then sequentially input into each residual module; after all residual modules are processed, they are sequentially input into each deconvolution module for processing;
[0011] The convolution module consists of a convolution layer, a BN layer, and an ELU activation function; the residual module consists of a residual layer, a BN layer, and an ELU activation function; the deconvolution module consists of a deconvolution layer, a BN layer, and an ELU activation function;
[0012] The ELU activation function formula is:
[0013]
[0014] Where x is the input value and α is the hyperparameter that controls the negative input value.
[0015] Preferably, after the convolution layer in the first convolution module in the generator network performs feature extraction on the input image to generate a feature map, an attention mechanism is used for processing; the attention mechanism includes pooling, full connection, activation function, full connection, and Sigmoid process. First, the pooling process is used for the feature map with a size of H*W*C to generate a 1*1*C vector, where H represents height, W represents width, and C represents the number of channels; the 1*1*C vector is input into the fully connected layer to reduce the dimension to obtain a 1*1*(C / 4) vector, and the activation function is used to increase the dimension to a 1*1*C vector, and Sigmoid is used to obtain a 1*1*C vector. Finally, the channel weighted method is used to obtain the feature map processed by the attention mechanism.
[0016] Preferably, the discriminator network includes a plurality of convolution modules, a feature elimination module, and a classification module;
[0017] The convolution module uses feature fusion to generate a multi-level fused feature map; the feature elimination module structure is composed of a mask matrix, which contains digital elements of 0 and 1. The mask matrix has the same size as the multi-level feature map. The multi-level fused feature map is input into the feature elimination module to eliminate the area with a corresponding position of 0 in the multi-level feature map to obtain the last-level feature map; the classification module determines the probability that the input image belongs to a real high-resolution image based on the last-level feature map.
[0018] Preferably, the convolution module adopts an upsampling multi-level fusion method to generate a multi-level fusion feature map.
[0019] Preferably, the weld X-ray image enhancement model is trained using the data set, the high-resolution weld X-ray image label is set to 1, the generated image label is set to 0, the generator network and the discriminator network are trained separately and alternately, and the discriminator network is trained first. 1 times, train the generator network once, and repeat this training process. 1 The value range is 2 to 5.
[0020] Preferably, a generator network loss function is constructed, and the generator network is trained using the data set;
[0021] The generator network loss function uses cross entropy to calculate the generation category loss, and adds image generation loss to measure the loss of the difference between the generated image and the high-resolution image; the generator network loss function is:
[0022]
[0023] Among them, G represents the generator, D represents the discriminator, E represents the expected loss value, and x L represents a low-resolution image, x H represents a high-resolution image, G(x L ) represents the sample generated by the generator based on the low-resolution image, D(G(x L ) represents the probability of the discriminator determining whether the generated image is the corresponding high-resolution image, and n is the number of samples;
[0024] One round of iterative training of the generator network includes the forward propagation and back propagation processes of the network. After each round of iterative training of the generator network is completed, the generated image and the high-resolution image are input into the generator network loss function. After calculating the generator network loss value, the next round of training process is restarted.
[0025] Preferably, a discriminator network loss function is constructed, and the discriminator network is trained using the data set;
[0026] The discriminator network loss function formula is:
[0027] D loss =Ex L [log(1-D(G(x L )]+Ex H [log(D(x H )]
[0028] Among them, G represents the generator, D represents the discriminator, E represents the expected loss value, and x L represents a low-resolution image, x H represents a high-resolution image, G(x L ) represents the sample generated by the generator based on the low-resolution image, D(G(x L ) represents the probability of the discriminator determining whether the generated image is the corresponding high-resolution image, D(x H ) represents the probability of the discriminator determining whether the input high-resolution image is a high-resolution image;
[0029] One round of iterative training of the discriminator network includes the forward propagation and back propagation processes of the network. After each round of iterative training of the discriminator network is completed, the generated image and the high-resolution image are input into the discriminator network loss function. After calculating the discriminator network loss value, the next round of training process is restarted.
[0030] Preferably, a random seed method is used to divide the data sets of high-resolution weld X-ray images and low-resolution weld X-ray images containing corresponding relationships into a training set, a validation set, and a test set, with a division ratio of 8:1:1, and the weld X-ray image enhancement model is trained.
[0031] Preferably, each high-resolution weld X-ray image in the sample is preprocessed to obtain a low-resolution weld X-ray image, and the preprocessing method is: downsampling the high-resolution weld X-ray image, then adding Gaussian noise, and finally performing mean blurring.
[0032] The advantages of the present invention compared with the prior art are:
[0033] (1) The present invention constructs a generator network consisting of a convolution module, a residual module, and a deconvolution module; the convolution layer in the convolution module extracts features from the input image and generates a feature map, and then uses an attention mechanism to process the feature map, so that the generator network assigns higher weights to key feature information, and the network as a whole pays more attention to the key feature information.
[0034] (2) The present invention adds a loss value for measuring the difference between the generated image and the high-resolution image to the generator network loss function. By adding this loss value, the difference between the generated image and the high-resolution image is more accurately described, and the overall network training process pays more attention to the difference between the generated image and the high-resolution image.
[0035] (3) The present invention constructs a discriminator network consisting of a convolution module, a feature elimination module, and a classification module; the feature map generated by the convolution module adopts a feature fusion method to generate a multi-level fusion feature map, so that the discriminator network extracts feature information of different scales and enhances the overall feature extraction capability of the network; the feature elimination module adopts a mask matrix to randomly eliminate certain position areas in the feature map, thereby strengthening the expression of each feature area and avoiding the overfitting problem caused by multiple trainings.
[0036] (4) The present invention adopts multiple methods such as downsampling, adding Gaussian noise, and mean blurring to obtain low-resolution weld X-ray images, and constructs a large number of data sets containing corresponding high-resolution weld X-ray images and low-resolution weld X-ray images. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0038] Figure 1 This is a flow chart of a weld X-ray image enhancement method based on a generative adversarial network according to an embodiment of the present invention;
[0039] Figure 2 Generate an adversarial network schematic diagram for an embodiment of the present invention;
[0040] Figure 3 This is a schematic diagram of a weld X-ray image enhancement model based on a generative adversarial network according to an embodiment of the present invention;
[0041] Figure 4 This is a schematic diagram of the attention mechanism of an embodiment of the present invention;
[0042] Figure 5 This is a flow chart of a method for fusing feature maps generated by five convolutional modules in a discriminator network according to an embodiment of the present invention;
[0043] Figure 6 The figure is a flow chart of feature elimination according to an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0045] The present invention proposes a weld X-ray image enhancement method based on a generative adversarial network. Figure 1 As shown, the specific steps include:
[0046] Step 1: Use high-precision X-ray imaging equipment to obtain high-resolution weld X-ray images, where the X-ray source in the high-precision X-ray imaging equipment is 100KV voltage, 250mA current, and the focal spot size is 5 microns, and the spatial resolution of the high-resolution image is 1 to 20 microns. Preprocess the high-resolution weld X-ray image to obtain a low-resolution weld X-ray image, and the spatial resolution of the low-resolution image is 200 to 400 microns, and construct a high-resolution weld X-ray image and low-resolution weld X-ray image dataset containing corresponding relationships.
[0047] In this embodiment, the high-resolution weld X-ray image is a 16-bit single-channel grayscale image, with a total of 5810 images and an image size of 1024*1024. The preprocessing method is:
[0048] The high-resolution weld X-ray image downsampling uses bilinear interpolation to generate 5810 low-resolution weld X-ray images. The pixel calculation formula after bilinear interpolation is as follows:
[0049] f(x+u,y+v)=(1-u)(1-v)f(x,y)+(1-u)vf(x,y+1)+u(1-v)f(x+1,y)+uvf(x+1,y+1)
[0050] Where f(x,y) represents the pixel value of point (x,y) in the original image, f(x+u,y+v) represents the pixel value of point (x+u,y+v) in the image, and u and v represent the decimal part of the floating-point coordinates in the interval [0,1).
[0051] Gaussian noise is added to the high-resolution weld X-ray image to generate 5810 low-resolution weld X-ray images. The Gaussian noise formula is as follows:
[0052]
[0053] Where n(x, y) represents the Gaussian noise value added to the point (x, y) in the image, and δ represents the standard deviation of the Gaussian distribution.
[0054] The mean fuzzy formula of high-resolution weld X-ray images is as follows:
[0055]
[0056] Where f(x,y) represents the pixel value of the (x,y) point after the image is mean blurred, and k represents the convolution kernel size. In this embodiment, k=3, and 5810 low-resolution weld X-ray images are generated; k=5 again, and 5810 low-resolution weld X-ray images are generated.
[0057] In summary, 5810 high-resolution weld X-ray images were preprocessed to obtain low-resolution weld X-ray images, and a total of 23240 low-resolution weld X-ray images were obtained.
[0058] Step 2: Build a weld X-ray image enhancement model based on generative adversarial network
[0059] Generative adversarial networks consist of a generator network and a discriminator network, such as Figure 2 As shown, the input of the generator network is a low-resolution weld X-ray image, and the output is a generated image; the input of the discriminator network is a high-resolution weld X-ray image and a generated image, and the output is the discrimination result that this image is a real weld X-ray image.
[0060] The generator network consists of 3 convolution modules, 2 residual modules, and 3 deconvolution modules, such as Figure 3 As shown in the figure, the network convolution module consists of a convolution layer, a BN layer, and an ELU activation function; the residual module consists of a residual layer, a BN layer, and an ELU activation function; and the deconvolution module consists of a deconvolution layer, a BN layer, and an ELU activation function.
[0061] The ELU activation function formula is as follows:
[0062]
[0063] In this embodiment, the value of α is set to 0.1.
[0064] After the convolution layer in the first convolution module in the generator network extracts features from the input image to generate a feature map, it is processed using the attention mechanism. The attention mechanism includes pooling, full connection, activation function, full connection, and Sigmoid process, such as Figure 4 As shown in the figure, firstly, the pooling process is used for the feature map of size H*W*C to generate a 1*1*C vector; the 1*1*C vector is input into the fully connected layer to reduce the dimension to obtain a 1*1*(C / 4) vector, and then the activation function is used to increase the dimension to a 1*1*C vector, and Sigmoid is used to obtain a 1*1*C vector. Finally, the channel weighted method is used to obtain the feature map processed by the attention mechanism, as shown in Figure 4shown.
[0065] The discriminator network consists of 5 convolution modules, 1 feature elimination module, and 1 classification module, such as Figure 5 As shown in the figure, the convolution module structure consists of a convolution layer, a BN layer, and an ELU activation function; the feature elimination module structure consists of a mask matrix, which contains 0 and 1 digital elements; the classification module consists of a fully connected layer, an ELU activation function, a fully connected layer, and a Sigmoid function.
[0066] The feature maps generated by the five convolution modules in the discriminator network are fused by feature fusion to generate a multi-level fusion feature map; the multi-level fusion feature map is input into the feature elimination module to obtain the final feature map. The feature fusion method of the feature maps generated by the five convolution modules in the discriminator network adopts the upsampling multi-level fusion method. The fusion method is as follows: Figure 5 As shown, the feature map F5 generated by the convolution module 5 is upsampled and fused with the feature map F4 generated by the convolution module 4 to generate the feature map P4. The feature map P4 is upsampled and fused with the feature map F3 generated by the convolution module 3 to generate the feature map P3; the feature map P3 is upsampled and fused with the feature map F2 generated by the convolution module 2 to generate the feature map P2; the feature map P2, the feature map P3, the feature map P4, and the feature map F5 are horizontally connected to form a multi-level feature map.
[0067] The structure of the feature elimination module is composed of a mask matrix, which has the same size as the multi-level feature map. The mask matrix randomly generates numbers 0 and 1, and the ratio of 0 to 1 is 2:98. After the multi-level feature map is input into the feature elimination module, the area with the corresponding position of 0 in the multi-level feature map is eliminated, such as Figure 6 shown.
[0068] Step 3: Train the weld X-ray image enhancement model to obtain a trained weld X-ray image enhancement model, and the training data set uses high-resolution weld X-ray images and low-resolution weld X-ray images containing corresponding relationships.
[0069] Using the random seed method, the high-resolution weld X-ray image dataset and the low-resolution weld X-ray image dataset containing corresponding relationships are divided into training set, validation set, and test set, with a division ratio of 8:1:1, that is, the number of high-resolution-low-resolution image pairs in the training set, validation set, and test set are 18529, 2324, and 2324, respectively.
[0070] In order to train the generator network and the discriminator network, the generator network loss function and the discriminator network loss function are constructed respectively.
[0071] The generator network loss function uses cross entropy to calculate the generated category loss, and adds image generation loss to measure the loss of the difference between the generated image and the high-resolution image. The generator network loss function formula is as follows:
[0072]
[0073] Among them, G represents the generator, D represents the discriminator, E represents the expected loss value, and x L represents a low-resolution image, x H represents a high-resolution image, G(x L ) represents the sample generated by the generator based on the low-resolution image, D(G(x L )) represents the judgment probability of the discriminator for the low-resolution image, and n is the number of samples.
[0074] The discriminator network loss function formula is as follows:
[0075] D loss =Ex L [log(1-D(G(x L )]+Ex H [log(D(x H )]
[0076] In this embodiment, the training process of the generative adversarial network model uses the NVIDIA RXT4090 graphics card, the deep learning framework is pytorch1.8 version, and the python3.9 language is used; the reasoning process uses the NVIDIA RXT4060 graphics card, the deep learning framework is pytorch1.8 version, and the python3.9 language is used. Set the high-resolution weld X-ray image label to 1, and the generated weld X-ray image label to 0. The generator network and the discriminator network are trained separately and alternately. First, the discriminator network is trained 3 times, and then the generator network is trained once, and this training process is repeated in sequence. The discriminator network is trained 30,000 times and the discriminator network is trained 10,000 times.
[0077] Step 4: Input the low-resolution weld X-ray image that needs to be enhanced into the trained weld X-ray image enhancement model to obtain an enhanced weld X-ray image.
[0078] The above-described embodiments are only preferred specific implementations of the present invention. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.
Claims
1. A weld X-ray image enhancement method based on generative adversarial network, characterized in that: include: Collect high-resolution weld X-ray image samples with a spatial resolution of 1 to 20 microns, pre-process each high-resolution weld X-ray image in the sample to obtain a low-resolution weld X-ray image, where the spatial resolution of the low-resolution weld X-ray image is 200 to 400 microns; construct a data set of high-resolution weld X-ray images and low-resolution weld X-ray images containing corresponding relationships; A weld X-ray image enhancement model based on generative adversarial network is constructed, in which the generator network input is a low-resolution weld X-ray image and the output is a generated image; the discriminator network input is a high-resolution weld X-ray image and the generated image, and the output is the discriminant probability of whether the generated image is a high-resolution weld X-ray image; The data set is used to train the weld X-ray image enhancement model; the low-resolution weld X-ray image that needs to be enhanced in image quality is input into the trained weld X-ray image enhancement model, and the generated image output by the generator network is the enhanced weld X-ray image.
2. The weld X-ray image enhancement method based on generative adversarial network according to claim 1, characterized in that: The generator network includes multiple convolution modules, multiple residual modules, and multiple deconvolution modules; after the low-resolution weld X-ray image is processed by each convolution module, it is sequentially input into each residual module; after all residual modules are processed, it is sequentially input into each deconvolution module for processing; The convolution module consists of a convolution layer, a BN layer, and an ELU activation function; the residual module consists of a residual layer, a BN layer, and an ELU activation function; the deconvolution module consists of a deconvolution layer, a BN layer, and an ELU activation function; The ELU activation function formula is: Where x is the input value and α is the hyperparameter that controls the negative input value.
3. The weld X-ray image enhancement method based on generative adversarial network according to claim 2, characterized in that: After the convolution layer in the first convolution module in the generator network extracts features from the input image to generate a feature map, it is processed using the attention mechanism; the attention mechanism includes pooling, full connection, activation function, full connection, and Sigmoid process. First, the pooling process is used for the feature map of size H*W*C to generate a 1*1*C vector, where H represents height, W represents width, and C represents the number of channels; the 1*1*C vector is input into the fully connected layer to reduce the dimension to obtain a 1*1*(C / 4) vector, and the activation function is used to increase the dimension to a 1*1*C vector, and Sigmoid is used to obtain a 1*1*C vector. Finally, the channel weighted method is used to obtain the feature map processed by the attention mechanism.
4. The weld X-ray image enhancement method based on generative adversarial network according to claim 1, characterized in that: The discriminator network includes multiple convolution modules, a feature elimination module, and a classification module; The convolution module uses feature fusion to generate a multi-level fusion feature map. The feature elimination module consists of a mask matrix, which contains 0 and 1 digital elements. The mask matrix has the same size as the multi-level feature map. The multi-level fusion feature map is input into the feature elimination module to eliminate the area with a corresponding position of 0 in the multi-level feature map to obtain the last-level feature map. The classification module determines the probability that the input image belongs to a real high-resolution image based on the feature map of the last layer.
5. The weld X-ray image enhancement method based on generative adversarial network according to claim 4, characterized in that: The convolution module adopts the upsampling multi-level fusion method to generate a multi-level fusion feature map.
6. The weld X-ray image enhancement method based on generative adversarial network according to claim 1, characterized in that: The weld X-ray image enhancement model is trained using the data set. The high-resolution weld X-ray image label is set to 1, the generated image label is set to 0, and the generator network and the discriminator network are trained separately and alternately. First, the discriminator network is trained m1 times, and then the generator network is trained 1 time. This training process is repeated in sequence, and m1 takes a value of 2 to 5.
7. The weld X-ray image enhancement method based on generative adversarial network according to claim 6, characterized in that: Construct the generator network loss function and use the dataset to train the generator network; The generator network loss function uses cross entropy to calculate the generation category loss, and adds image generation loss to measure the loss of the difference between the generated image and the high-resolution image; the generator network loss function is: Among them, G represents the generator, D represents the discriminator, E represents the expected loss value, and x L represents a low-resolution image, x H represents a high-resolution image, G(x L ) represents the sample generated by the generator based on the low-resolution image, D(G(x L ) represents the probability of the discriminator determining whether the generated image is the corresponding high-resolution image, and n is the number of samples; One round of iterative training of the generator network includes the forward propagation and back propagation processes of the network. After each round of iterative training of the generator network is completed, the generated image and the high-resolution image are input into the generator network loss function. After calculating the generator network loss value, the next round of training process is restarted.
8. The weld X-ray image enhancement method based on generative adversarial network according to claim 6, characterized in that: Constructing a discriminator network loss function, and using the data set to train the discriminator network; The discriminator network loss function formula is: D loss =Ex L [log(1-D(G(x L )]+Ex H [log(D(x H )] Among them, G represents the generator, D represents the discriminator, E represents the expected loss value, and x L represents a low-resolution image, x H represents a high-resolution image, G(x L ) represents the sample generated by the generator based on the low-resolution image, D(G(x L ) represents the probability of the discriminator determining whether the generated image is the corresponding high-resolution image, D(x H ) represents the probability of the discriminator determining whether the input high-resolution image is a high-resolution image; One round of iterative training of the discriminator network includes the forward propagation and back propagation processes of the network. After each round of iterative training of the discriminator network is completed, the generated image and the high-resolution image are input into the discriminator network loss function. After calculating the discriminator network loss value, the next round of training process is restarted.
9. The weld X-ray image enhancement method based on generative adversarial network according to claim 1, characterized in that: The random seed method is used to divide the data sets of high-resolution weld X-ray images and low-resolution weld X-ray images containing corresponding relationships into training set, validation set, and test set with a division ratio of 8:1:1, and the weld X-ray image enhancement model is trained.
10. The weld X-ray image enhancement method based on generative adversarial network according to claim 1, characterized in that: Each high-resolution weld X-ray image in the sample is preprocessed to obtain a low-resolution weld X-ray image. The preprocessing method is: downsampling the high-resolution weld X-ray image, adding Gaussian noise, and finally performing mean blurring.
Citation Information
Patent Citations
Welding seam X-ray photograph defect detection method
CN114187256A
X-ray weld defect image reconstruction method based on improved SRGAN
CN118314018A
Multi-feature fusion echo cancellation method and system based on self-attention transform network
WO2023044961A1
Cited By
X-ray multi-contrast imaging image quality enhancement method based on feature prior learning
CN119624813A
Image quality enhancement method for X-ray multi-contrast imaging based on feature prior learning
CN119624813B
B ultrasonic image enhancement processing method based on deep learning
CN119741225A
