A method for enhancing image quality in gamma-photon industrial reconstruction based on parallel deformable convolutional networks.

Abstract

This invention discloses a method for enhancing the quality of gamma-photon industrial reconstruction images based on parallel deformable convolutional networks, comprising the following steps: A gamma-photon industrial detection image containing complex noise is input into a constructed denoising model, which consists of convolutional modules and an adaptive feature fusion module. The convolutional modules are an initial convolutional block, a dilated convolutional block, and a deformable convolutional block. These three convolutional blocks are established in parallel, utilizing the different receptive field sizes of the dilated and deformable convolutional blocks to fully capture global feature information of the gamma-photon industrial image, while the initial convolutional block captures local feature information. The adaptive feature fusion module then combines local and global feature information to effectively handle the differences in noise and texture details in the gamma-photon industrial image. This outputs a high-quality reconstructed image, improving the imaging capabilities of gamma-photon industrial detection.

Description

Technical Field

[0001] This invention relates to the fields of image processing technology and gamma photon industrial image reconstruction, and particularly to a method for enhancing the quality of gamma photon industrial reconstructed images based on parallel deformable convolutional networks. Background Technology

[0002] With the rapid development of computer and image processing technologies, positron emission tomography (PET) has become an important non-destructive testing technique in the industrial field. PET technology captures high-energy gamma photons generated by positron annihilation to perform internal imaging of industrial objects, and is widely used in aerospace, metal material inspection, and other fields. However, in this application scenario, the quality of the reconstructed image is easily affected by complex noise.

[0003] Traditional image filters such as mean filtering and median filtering can remove noise to some extent, but their performance is less than ideal when dealing with complex boundary information. Mean filtering is suitable for removing additive Gaussian noise by averaging the gray values ​​of neighboring pixels. However, it blurs edge and detail information, leading to signal distortion. Median filtering excels at removing salt-and-pepper noise by replacing pixel values ​​with the median of neighboring pixels, eliminating extreme noise. However, it can still cause signal distortion when processing periodic signals and boundary information. Noise in gamma-photon industrial images is caused by scattering of gamma photons through metallic materials and physical decay effects. Traditional denoising techniques cannot effectively address these challenges and struggle to meet the demands of high-precision non-destructive testing.

[0004] Convolutional Neural Networks (CNNs) have been widely used in image denoising tasks in recent years due to their powerful learning capabilities and flexibility. Through convolution operations, CNNs can extract complex noise features from images. However, in the reconstruction of industrial large-axis spatial PET images, existing CNN models still have several problems: (1) In complex industrial noise environments, existing CNN models are unable to effectively handle high-intensity noise, resulting in unsatisfactory denoising effects. (2) Deep networks are prone to gradient vanishing, affecting training performance. (3) Simply expanding the receptive field can capture more feature information, but it increases the number of model parameters, leading to slower training. Therefore, how to design a denoising network suitable for γ-photon industrial image reconstruction has become an urgent research problem to be solved. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for enhancing the quality of industrial reconstructed images based on deformable convolutional networks using gamma photons.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a method for enhancing the quality of gamma-photon industrial reconstruction images based on parallel deformable convolutional networks, comprising the following steps:

[0007] Step 1: Use the original γ-photon industrial reconstructed image obtained by the ordered subset maximum expectation method OSEM as input;

[0008] Step 2: Learn global and local features of the industrial reconstructed image through the convolution module;

[0009] 1) The convolution module mainly consists of three parallel network branches. The first layer of the three network branches consists of an initial convolutional kernel, a dilated convolutional kernel, a deformable convolutional kernel, and an activation function ReLU, respectively, to generate initial global and local feature information maps.

[0010] 2) The second to fifth layers consist of an initial convolutional kernel, an expanded convolutional kernel, a deformable convolutional kernel, an activation function, and batch normalization, respectively, to improve the nonlinear expression capability of the network branches and obtain more refined and complex noise and texture detail feature maps;

[0011] 3) Each network branch is finally processed by 32 convolutional kernels of size 3×3 to output a feature information map;

[0012] Step 3: Input global and local feature information into the adaptive feature fusion module. The adaptive feature fusion module contains a global feature fusion block and a local feature fusion block, which fuses the differences in noise and texture details in the reconstructed image from different regions to improve the model's ability to recover the details of the reconstructed image, as detailed below:

[0013] 1) The local feature information F1 from the global and local feature information output by the convolution module is fused by the attention of the global feature fusion module to obtain the fused global feature information map F′2. The operation is as follows:

[0014] After linearly mapping the global feature information F2, multiply it with the local feature information F1; then obtain the corresponding weight ω through activation function and normalization. F The corresponding weight ω F After weighted summation with local feature information F1, a new feature information map F′1 is obtained. Then, the new feature information map is linearly mapped and added to the global feature information map F2 to obtain the fused global feature information map F′2.

[0015] 2) The local feature information F2 from the global and local feature information output by the convolution module is fused by the attention of the local feature fusion module to obtain the fused local feature information map F′1. The operation is as follows:

[0016] After linearly mapping the local feature information F1, it is multiplied by the global feature information F2; then, the corresponding weight ω is obtained through activation function and normalization. F ′, corresponding weight ω FThe new feature information map F′2 is obtained by weighted summation of F′ and global feature information F2. Then, the new feature information map is linearly mapped and added to the local feature information map F1 to obtain the fused local feature information map F′1.

[0017] Step 4: The output of the adaptive feature fusion module is then subjected to a 1×1 vector convolution operation (Conv) to obtain a high-quality γ-photon industrial reconstructed image;

[0018] As a preferred embodiment, the initial convolutional kernel in the three parallel network branches is the first network branch to learn local feature information of the input image; the deformable convolutional kernel is the second network branch, which is directional, to learn global feature information of the input image; and the dilated convolutional kernel is the third network branch, which allows the convolutional kernel to sample on different receptive fields, thereby capturing multi-scale feature information.

[0019] As a preferred embodiment, the deformable convolution kernel is obtained by performing Fourier interpolation on the convolution kernel. First, a Fourier transform is performed on the convolution kernel. Let the convolution kernel be K(m,n), then the Fourier transform is performed as follows:

[0020]

[0021] Where x and y are the sizes of the convolution kernel, and F(K)(μ, v) represents the convolution kernel in the frequency domain;

[0022] Interpolation of the convolution kernel in the frequency domain is performed as follows:

[0023]

[0024] Where δ is the deformation strength;

[0025] Finally, an inverse Fourier transform is performed to obtain the deformable convolution kernel K′(m, n), as shown below:

[0026]

[0027] As a preferred embodiment, the dilation rate of the expanded convolution kernel is 2 or 7.

[0028] The beneficial effects of this invention are as follows: By introducing a parallel three-branch network, this method significantly improves denoising performance. Simultaneously, while keeping the training objects unchanged, it indirectly reduces the network depth, effectively avoiding gradient vanishing or exploding problems, reducing training costs, and enhancing the network's generalization ability. Without increasing computational load, it increases the receptive field, avoids information loss, and strengthens the ability to deeply mine multi-scale information. It can more flexibly fit complex geometric shapes in industrial part images, improving the network's generalization ability. Through deformable convolution to extract features, the denoised image contours and edges can better maintain a realistic effect.

[0029] Traditional convolutions, due to their fixed sampling methods, struggle to effectively fit complex geometric shapes. This invention overcomes the limitations of regular sampling by performing Fourier interpolation on ordinary convolution kernels, allowing the kernels to adaptively sample around their current position. This enhances the flexibility and accuracy of feature extraction, significantly improving the network's adaptability to input features.

[0030] The adaptive feature fusion module effectively combines global and local feature information, which can better handle noise and texture differences in different regions and significantly improve the model's ability to recover spatial detail information. Attached Figure Description

[0031] Figure 1 This is a flowchart of a gamma-ray industrial reconstruction image quality enhancement method.

[0032] Figure 2 This is a schematic diagram of a parallel deformable convolutional network structure.

[0033] Figure 3(a) is a schematic diagram of the global feature fusion block structure.

[0034] Figure 3(b) is a schematic diagram of the local feature fusion block structure.

[0035] Figure 4 This is a schematic diagram illustrating the denoising effect of an original industrial ring pipeline inspection image in a specific embodiment. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, a method for enhancing the quality of industrial reconstructed images based on deformable convolutional networks using gamma photons includes the following steps:

[0038] Step 1: Use the original gamma-photon industrial reconstructed image obtained by the OSEM algorithm as input;

[0039] Step 2: The convolutional module is used to learn the global and local features of the industrially reconstructed image. For example... Figure 2 As shown:

[0040] 1) The convolution module consists of three parallel network branches. The first layer of each of the three network branches consists of 16 initial convolutional kernels of size 3×3, 16 dilated convolutional kernels of size 2×2 with an inflation rate of 2, and 16 deformable convolutional kernels of size 3×3 and the activation function ReLU, generating initial global and local feature maps.

[0041] 2) The second to sixth layers consist of 16 initial convolutional kernels of size 3×3, 16 dilated convolutional kernels of size 2×2, with an inflation rate of 7, 16 deformable convolutional kernels of size 3×3, activation functions, and batch normalization, which improves the nonlinear expression ability of the network branches and obtains more refined and complex noise and texture detail feature maps.

[0042] Fourier interpolation is performed on deformable convolution kernels in parallel network branches. The interpolation steps are as follows: First, perform a Fourier transform on the convolution kernel. Let the convolution kernel be K(m,n), and perform the Fourier transform:

[0043]

[0044] Where x and y are the sizes of the convolution kernel, and f(K)(μ, v) represents the convolution kernel in the frequency domain.

[0045] Interpolation of the convolution kernel in the frequency domain is performed as follows:

[0046]

[0047] Where δ is the deformation strength

[0048] Finally, an inverse Fourier transform is performed to obtain the deformable convolution kernel K′(m, n), as shown below:

[0049]

[0050] 3) Each network branch is finally processed by 32 convolutional kernels of size 3×3 to output a feature information map.

[0051] Step 3: Input global and local feature information into the adaptive feature fusion module, which contains a global feature fusion block and a local feature fusion block. This module fuses the differences in noise and texture details in the reconstructed image from different regions, improving the model's ability to recover details from the reconstructed image.

[0052] 1) As shown in Figure 3(a), the local feature information F1 in the global and local feature information output by the convolution module is fused by the attention of the global feature fusion module to obtain the fused global feature information map F′2.

[0053] 2) As shown in Figure 3(b), the local feature information F2 from the global and local feature information output by the convolution module is fused by the attention of the local feature fusion module to obtain the fused local feature information map F′1.

[0054] In the global feature fusion block, the local feature information F1 from the input adaptive feature fusion module is integrated into the global feature information F2. The integration process is as follows:

[0055] After linearly mapping the global feature information F2, it is multiplied by the local feature information F1. Then, the corresponding weights ω are obtained through activation functions and normalization. F The corresponding weight ω F After weighted summation with local feature information F1, a new feature information map F′1 is obtained. Then, the new feature information map is linearly mapped and added to the global feature information map F2 to obtain the fused global feature information map F′2.

[0056] In the local feature fusion block, the global feature information F2 input to the adaptive feature fusion module is integrated into the local feature information F1. The integration process is as follows:

[0057] After linearly mapping the local feature information F1, it is multiplied by the global feature information F2. Then, the corresponding weights ω are obtained through activation functions and normalization. F ′, corresponding weight ω F The new feature information map F′2 is obtained by weighted summation of the global feature information F2. Then, the new feature information map is linearly mapped and added to the local feature information map F1 to obtain the fused local feature information map F′1.

[0058] Step 4: The output of the adaptive feature fusion module is then processed by a 1×1 Conv to obtain a high-quality γ-photon industrial reconstruction image.

[0059] In a specific example of annular pipe inspection imaging implementation, such as Figure 4 As shown, Figure (a) is the original slice of the reconstructed image of the circular pipe, Figure (b) is the denoised slice of the classic DnCNN method, and Figure (c) is the denoised slice of the method of the present invention. It can be seen that the method of the present invention enhances the detail information of the original reconstructed image while removing noise.

[0060] The above embodiments are merely illustrative of the principles and effects of the present invention, as well as some examples of its application, and are not intended to limit the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention.

Claims

1. A method for enhancing the quality of industrial reconstructed gamma-photon images based on parallel deformable convolutional networks, comprising the following steps: Step 1: Use the original γ-photon industrial reconstructed image obtained by the ordered subset maximum expectation method OSEM as input; Step 2: Learn global and local features of the industrial reconstructed image through the convolution module; 1) The convolution module mainly consists of three parallel network branches. The first layer of the three network branches consists of an initial convolutional kernel, a dilated convolutional kernel, a deformable convolutional kernel, and an activation function ReLU, respectively, to generate initial global and local feature information maps. 2) The second to fifth layers consist of an initial convolutional kernel, an expanded convolutional kernel, a deformable convolutional kernel, an activation function, and batch normalization, respectively, to improve the nonlinear expression capability of the network branches and obtain more refined and complex noise and texture detail feature maps; 3) Each network branch is finally processed by 32 convolutional kernels of size 3×3 to output a feature information map; Step 3: Input global and local feature information into the adaptive feature fusion module. The adaptive feature fusion module contains a global feature fusion block and a local feature fusion block, which fuses the differences in noise and texture details in the reconstructed image from different regions to improve the model's ability to recover the details of the reconstructed image, as detailed below: 1) The local feature information F1 from the global and local feature information output by the convolution module is fused by the attention of the global feature fusion module to obtain the fused global feature information map F′2. The operation is as follows: After linearly mapping the global feature information F2, multiply it with the local feature information F1; then obtain the corresponding weight ω through activation function and normalization. F The corresponding weight ω F After weighted summation with local feature information F1, a new feature information map F′1 is obtained. Then, the new feature information map is linearly mapped and added to the global feature information map F2 to obtain the fused global feature information map F′2. 2) The local feature information F2 from the global and local feature information output by the convolution module is fused by the attention of the local feature fusion module to obtain the fused local feature information map F′1. The operation is as follows: After linearly mapping the local feature information F1, it is multiplied by the global feature information F2; then the corresponding weight ω is obtained through activation function and normalization. F ′, corresponding weight ω F The new feature information map F′2 is obtained by weighted summation of F′ and global feature information F2. Then, the new feature information map is linearly mapped and added to the local feature information map F1 to obtain the fused local feature information map F′1. Step 4: The output of the adaptive feature fusion module is then subjected to a 1×1 vector convolution operation (Conv) to obtain a high-quality gamma-photon industrial reconstructed image.

2. The method for enhancing the quality of gamma-photon industrial reconstruction images based on parallel deformable convolutional networks according to claim 1, characterized in that, The initial convolutional kernel in the three parallel network branches forms the first network branch to learn local feature information of the input image; the deformable convolutional kernel forms the second network branch, and the deformable convolutional kernel is directional to learn global feature information of the input image. The dilated convolutional kernel is the third branch of the network, which allows the convolutional kernel to sample on different receptive fields, thereby capturing multi-scale feature information.

3. The method for enhancing the quality of gamma-photon industrial reconstruction images based on parallel deformable convolutional networks according to claim 2, characterized in that, The deformable convolution kernel is obtained by performing Fourier interpolation on the convolution kernel. First, a Fourier transform is performed on the convolution kernel; let the convolution kernel be K(m,n), and then perform the Fourier transform: Where x and y are the sizes of the convolution kernel, and f(K)(μ, v) represents the convolution kernel in the frequency domain; Interpolation of the convolution kernel in the frequency domain is performed as follows: Where δ is the deformation strength; Finally, an inverse Fourier transform is performed to obtain the deformable convolution kernel K′(m, n), as shown below:

4. The method for enhancing the quality of gamma-photon industrial reconstruction images based on parallel deformable convolutional networks according to claim 2, characterized in that, The expansion rate of the expanded convolution kernel is 2 or 7.

Images