Image noise removal method based on linear array X-ray camera
Through bright field correction, adaptive non-uniformity compensation and deep learning generative adversarial networks, the noise problem in linear array X-ray camera imaging is solved, the imaging quality is improved and artifacts are reduced, making it suitable for non-destructive testing and security inspection.
Patent Information
- Application Number
- CN202510966379.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-10-17
AI Technical Summary
There is stripe noise in the pixel and time directions in the imaging of linear array X-ray cameras, which affects the imaging quality and subsequent image analysis.
Bright field correction and adaptive non-uniformity compensation algorithms are used to process pixel-wise noise. The deep learning U-Net++ framework and generative adversarial networks are combined to combat noise. Line tracing and Lagrangian interpolation are used to repair noise. Dynamic integration and neural network correction are performed on temporal noise. A generative adversarial network is used to remove ring artifacts in three-dimensional images.
It effectively eliminates streak noise from linear array X-ray cameras, improves the quality of two-dimensional images, and reduces ring artifacts in three-dimensional images. It is suitable for non-destructive testing of large metal workpieces and security inspection imaging.
Smart Images
Figure CN120807346A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of image processing, and particularly relates to an image noise removal method based on a linear array X-ray camera. BACKGROUND
[0002] Computer tomography is a technology of penetrating an object by X-rays or gamma rays and receiving the attenuation signals by a detector, and generating the cross-sectional images of the object inside through computer processing and reconstruction. It has been widely used in materials, aerospace, mechanical manufacturing and other aspects by virtue of its detection of internal defects of workpiece equipment without damaging the workpiece equipment.
[0003] An electron induction accelerator is a device for promoting electrons to do circular acceleration by using a magnetic field. A metal target is placed at the end of the electron motion trajectory inside the accelerator, and X-rays can be generated by bombarding the target material with accelerated electrons. Compared with a linear electron accelerator X-ray source, it can increase the speed of electrons to a height that cannot be reached by linear acceleration under the condition of the same device volume, thereby generating higher energy X-rays. However, the use of the circular acceleration method causes more electron loss, resulting in the disadvantages of low exit dose and large dose fluctuation, which causes strip noise in the pixel direction and the time direction of imaging.
[0004] In the pixel direction, the relative positions of each pixel point and the target point of the electron induction accelerator are different. The X-rays generated by the bombardment of electrons have an angular distribution, and the dose levels are different at different spatial angles. As a fan-shaped source, the distance from each pixel of the vertically placed camera to the target point is not the same, resulting in different X-ray attenuations.
[0005] In the time direction, the dose of the electron accelerator itself fluctuates at the pixel level, resulting in inconsistent dose emission at each time.
[0006] The existence of these noises seriously affects the imaging quality of the linear array X-ray camera and brings great trouble to subsequent image analysis and application. Therefore, it is of great practical significance to develop an effective linear array X-ray camera image noise removal method. SUMMARY
[0007] The application aims to provide an image noise removal method based on a linear array X-ray camera, which adopts a bright field correction method and a method based on an adaptive non-uniformity compensation algorithm to post-process the strip noise of the reconstructed image, including: establishing a pixel sensitivity reference through bright field image statistics, performing bright-dark field correction under no light and corrected light, and interpolating and repairing the neighborhood of bad points; developing a GAN three-dimensional image noise removal system based on the U-Net++ framework by using deep learning.
[0008] To achieve the above-mentioned purpose, the application provides the following technical scheme:
[0009] An image noise removal method based on a linear array X-ray camera, comprising the following steps:
[0010] 1. Obtain a two-dimensional blank image without a workpiece based on an electron induction accelerator, and process the noise in the time direction and the pixel direction respectively;
[0011] 2. Obtain projection values at different angles based on a rotating workpiece, and reconstruct a three-dimensional image;
[0012] 3. Establish an algorithm model based on deep learning to reduce the central ring-shaped artifacts of the CT image in the three-dimensional reconstructed image.
[0013] Preferably, the pixel direction noise is processed by using a bright field correction method, which specifically comprises:
[0014] First, obtain a bright field image matrix Calculate the mean value in the pixel direction
[0015] Then, divide the to-be-processed image matrix M t,p by row to obtain a pixel correction matrix
[0016] Preferably, the time direction noise is processed by using a correction method based on adaptive non-uniformity, which specifically comprises:
[0017] First, adjust the preliminary noise reduction by dynamic integration time, and subtract the background noise corresponding to the integration time in real time; then, input the neural network to correct the residual high-frequency noise.
[0018] Preferably, the is calculated by using an arithmetic mean method, and the formula is:
[0019] Preferably, the blank pixel region selection standard is that the number of continuous unobstructed pixels exceeds 10% of the total number of pixels.
[0020] Preferably, the method is suitable for an imaging system in which an electron induction accelerator is used as an X-ray source.
[0021] Preferably, the image after the time direction correction is used for three-dimensional tomographic reconstruction, and a deep learning algorithm can be further used to eliminate ring-shaped artifacts.
[0022] Preferably, a U-Net++ structure is used, a residual convolution module is embedded to capture the spatial continuity features of three-dimensional artifacts, a dynamic attention mechanism is used, a channel-spatial dual attention module is introduced in the decoding layer, and the weight of the artifact area is adaptively enhanced.
[0023] Preferably, the discriminator architecture adopts the PatchGAN discriminator, whose convolution kernel size is 5*5; and adopts a hybrid loss function, including adversarial loss, perceptual loss and three-dimensional structure similarity loss.
[0024] Compared with the prior art, the present invention provides an image noise removal method based on a linear array X-ray camera, which has the following beneficial effects:
[0025] The present invention utilizes a comprehensive framework of bidirectional streak noise correction and GAN joint optimization to achieve efficient denoising of linear array X-ray images, reducing noise while preserving image details.
[0026] Bidirectional correction is performed in both pixel and time directions. In the pixel direction, the line tracing method is used to quickly locate the starting point and boundary of the horizontal stripe noise. Lagrange interpolation is combined to repair the sudden change of radiation value and improve the denoising efficiency. In the time direction, a non-uniform local algorithm is introduced to dynamically adjust the gain and bias coefficients.
[0027] In the GAN architecture, the generator simulates complex noise distribution based on a deep residual network, adapts to the noise characteristics of multiple devices through transfer learning, and the multi-scale discriminator combines Wasserstein distance optimization to solve the problem of residual artifacts.
[0028] The present invention can effectively eliminate the stripe noise of the linear array camera in the pixel and time directions, significantly improving the imaging quality of two-dimensional images; it uses a generative adversarial network to reduce ring artifacts in the reconstructed three-dimensional images, and is suitable for high-energy X-ray application scenarios such as non-destructive testing of large metal workpieces and security inspection imaging. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Fig. 1 Flow chart of the method of the present invention.
[0030] Fig. 2 This is a schematic diagram of the process of removing noise from linear array X-ray images based on a generative adversarial network in the present invention. DETAILED DESCRIPTION
[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0032] The present invention provides Figs. 1-2 The method for removing image noise based on a linear array X-ray camera includes the following steps:
[0033] 1. Obtain a two-dimensional blank image without a workpiece based on an electron induction accelerator, and process noise in the time direction and the pixel direction respectively;
[0034] 2. Obtain projection values at different angles based on a rotating workpiece, and reconstruct a three-dimensional image;
[0035] 3. Establish an algorithm model based on deep learning to reduce the central ring artifact of the CT image in the three-dimensional reconstructed image.
[0036] The pixel direction noise is processed by using a bright field correction method, which specifically includes:
[0037] First, obtain a bright field image matrix Calculate the pixel direction mean
[0038] Then, divide the to-be-processed image matrix M t,p by row to obtain a pixel correction matrix
[0039] The time direction noise is processed by using a correction method based on adaptive non-uniformity, which specifically includes:
[0040] First, adjust the preliminary noise reduction by dynamic integration time, and subtract the background noise corresponding to the integration time in real time; then, access the neural network to correct the residual high-frequency noise.
[0041] The calculation adopts an arithmetic mean method, and the formula is:
[0042] The blank pixel area selection standard in the hollow is that the number of continuous unobstructed pixels exceeds 10% of the total number of pixels.
[0043] The method is applicable to an imaging system using an electron induction accelerator as an X-ray source.
[0044] The image corrected in the time direction can be further used to eliminate the ring artifact in the three-dimensional tomographic reconstruction in combination with a deep learning algorithm.
[0045] The U-Net++ structure is adopted, and a residual convolution module is embedded to capture the spatial continuity features of the three-dimensional artifact. A dynamic attention mechanism is adopted, and a channel-spatial dual attention module is introduced in the decoding layer to adaptively enhance the weight of the artifact area.
[0046] The discriminator architecture adopts a PatchGAN discriminator, and the convolution kernel size is 5*5; a hybrid loss function is adopted, including an adversarial loss, a perception loss, and a three-dimensional structural similarity loss.
[0047] The transverse and longitudinal stripe noise is a common problem in a linear array camera, and the two-dimensional pixel noise and time noise are removed through line tracking positioning, Lagrange interpolation repair and non-uniform local algorithm. In three-dimensional CT reconstruction, for the ring-shaped artifacts commonly existing in three-dimensional images, a generative adversarial network based on U-Net architecture is designed to realize the removal of linear array image noise.
[0048] Two-dimensional image noise removal
[0049] 1. Pixel direction correction
[0050] The core of the line tracking method is to find the mutation point along the image column direction, and the upper and lower difference values of each pixel point are checked from the starting point of the image. Among them, the radiation value difference threshold is set to 5%-8% of the average radiation value of the image, which can be dynamically adjusted according to the actual imaging noise intensity. If the radiation value of the pixel point and the adjacent point above exceeds the threshold, it is marked as a mutation point.
[0051] The upper and lower difference values of the five equally spaced points near the mutation point are calculated again, and the spacing of the equally spaced points is set to 3-5 pixel units. If the difference values of the five points all exceed the threshold, it is determined that there is an edge point of stripe noise in the region, and the start and end points of the noise are determined.
[0052] Further, after determining the noise position, two uncontaminated pixel points above and below the noise region are selected as references, and a Lagrange polynomial fitting curve is obtained. The order of the Lagrange polynomial is selected to be 3 to ensure the fitting accuracy while avoiding overfitting, and the reference row radiation value is calculated to complete the repair of the noise region.
[0053] 2. Time direction correction
[0054] Affected by factors such as linear array detector, temperature, dose angle distribution, etc., the time direction noise fluctuation is large in this case, and an adaptive non-uniformity compensation algorithm is used to dynamically adjust the gain and bias.
[0055] First, the update period of the background noise library is set to 10 minutes / time to ensure that the noise change of the device can be reflected in time. The best integration time is selected from the pre-stored background noise library according to the average brightness of the current scene and the detector temperature. The selection range of the integration time is 50ms-200ms, and the interval is 10ms. After automatically deducting the background noise corresponding to the integration time, the trained neural network is accessed to correct the residual high-frequency noise.
[0056] The neural network adopts a structure of 3 convolution layers and 2 fully connected layers, the convolution kernel size is 3*3, the activation function is selected as ReLU function, and is obtained by training a large amount of linear array X-ray image data, the training iteration number is 5000 times, and the batch size is 32.
[0057] Three-dimensional image ring artifact removal
[0058] Three-dimensional reconstruction CT image ring artifact is a common image noise, the present application adopts the generation of adversarial network GAN to realize its removal, which is composed of two "adversarial" neural networks - generator and discriminator.
[0059] The task of the generator is to "forge" noise-free images, trying to remove noise and output pseudo-clean noise; the discriminator discriminates whether the image is a real noise-free data set or generated. Through the adversarial training, the generator improves the forgery ability, and finally generates a denoising result that is difficult to distinguish from the real noise-free image.
[0060] The generator architecture adopts U-Net structure, using an encoding-decoding network. The encoder contains 4 convolution blocks, each of which is composed of 2 convolution layers and 1 maximum pooling layer, and the convolution kernel size of the convolution layer is 3*3, and the step is 1, and the pooling size of the maximum pooling layer is 2*2. Noise features are extracted through the convolution layer, and the decoder is symmetrical to the encoder, containing 4 deconvolution blocks, each of which is composed of 2 deconvolution layers, and the convolution kernel size of the deconvolution layer is 3*3, and the step is 2.
[0061] In the framework, transfer learning is introduced, which is first pre-trained on a synthetic data set. The data set is an X-ray image with artificial noise, and the noise intensity is set according to the distribution characteristics of the actual ring artifact. The added noise types include ring patterns of different intensities and sizes. The number of iterations for pre-training is 3000 times, and the learning rate is 0.001. Then the generator parameters are fine-tuned through a small amount of real data, and the number of iterations for fine-tuning is 1000 times, and the learning rate is 0.0001.
[0062] It should be noted that the discriminator of the present application adopts multi-scale discrimination, which uses a pre-trained network to check the overall structure of the image and focuses on the noise area through an axial attention mechanism. The discriminator is composed of 3 sub-discriminators of different scales, which discriminate 1 / 4, 1 / 2 and the original size of the image respectively. During training, the generator needs to optimize the Wasserstein distance and the 3D gradient difference loss, wherein the weight coefficient of the Wasserstein distance is set to 0.6, and the weight coefficient of the 3D gradient difference loss is set to 0.4.
[0063] Finally, it should be pointed out that: the above-mentioned only for the preferred embodiments of the present application, and not for limiting the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the protection scope of the present application.
Claims
1. A method for removing image noise based on a linear array X-ray camera, characterized in that: The following steps are involved:
1. Acquire a two-dimensional blank image without artifacts based on an electron beta accelerator and process the noise in the time direction and pixel direction respectively; 2. Obtain projection values at different angles based on the rotating workpiece and reconstruct a three-dimensional image; 3. Establish a deep learning-based algorithm model to reduce the central ring artifacts of CT images in three-dimensional reconstructed images.
2. The method according to claim 1, characterized in that The bright field correction method is used to process pixel noise, including: First, obtain the bright field image matrix Calculate pixel-wise mean Then, the image matrix M to be processed t,p Divide row by row Get pixel correction matrix 3. The method according to claim 1, characterized in that The time direction noise is processed by a correction method based on adaptive non-uniformity, which includes: First, preliminary noise reduction is performed by adjusting the dynamic integration time to deduct the background noise corresponding to the integration time in real time; Then, the neural network is connected to correct the residual high-frequency noise.
4. The method according to claim 2, wherein: The The calculation adopts the arithmetic mean method, and the formula is:
5. The method according to claim 3, wherein: The selection criterion for the blank pixel area is that the number of continuous unobstructed pixels exceeds 10% of the total number of pixels.
6. The method according to claim 1, wherein: The method is applicable to an imaging system using an electron beta accelerator as an X-ray source.
7. The method according to claim 1, wherein: When the time-corrected image is used for three-dimensional tomographic reconstruction, a deep learning algorithm can be combined to further eliminate ring artifacts.
8. According to the method of claim 1, a 3D ring artifact removal system based on a generative adversarial network is proposed, characterized in that: The U-Net++ architecture is used, embedding a residual convolution module to capture the spatial continuity characteristics of 3D artifacts. A dynamic attention mechanism is used, introducing a channel-spatial dual attention module at the decoding layer to adaptively enhance the weights of artifact regions.
9. The method according to claim 1, wherein: The discriminator architecture uses the PatchGAN discriminator with a convolution kernel size of 5*5; it adopts a hybrid loss function, including adversarial loss, perceptual loss and 3D structure similarity loss.