A method and apparatus for enhancing CT projection images
By eliminating distortion and warping in CT projection images through a neural network model, improving resolution and removing metal artifacts, the image quality problem of security inspection CT equipment under baggage speed requirements is solved, and efficient item detection is achieved.
Patent Information
- Application Number
- CN202111584983.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-22
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2041-12-22
AI Technical Summary
Under the requirement of high baggage throughput, the spatial resolution of CT projection images is low and easily distorted, which affects the efficiency of subsequent manual image interpretation and intelligent recognition.
The first neural network model is used to eliminate the distortion and warping of CT projection images, the second neural network model is used to improve the resolution, and the third neural network model is used to remove metal artifacts. The resolution of the high-resolution detector is used for image enhancement.
It improves the quality of CT projection images, enhances the ability to identify items, especially metal items, and improves security inspection efficiency.
Smart Images

Figure CN116342396B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of X-ray detection technology, and in particular to a method and apparatus for enhancing CT projection images. Background Technology
[0002] In routine security checks, X-ray inspection machines are currently the most widely used and primary security equipment, extensively used in public security, transportation, government, and large-scale event venues. In more demanding security settings such as civil aviation and customs, CT scanners have seen rapid adoption. CT scanners not only obtain three-dimensional morphological data of the inspected object but also further obtain electron density and effective atomic number data at various points, thus possessing a relatively accurate ability to distinguish the types of substances. Therefore, compared to ordinary inspection machines, CT scanners have a significant advantage in detecting drugs, explosives, and biological tissues.
[0003] In actual use of CT scanners for security checks, the spatial resolution of the CT projection images obtained by the security equipment is significantly lower than that of the images obtained by the item inspection machine due to the speed requirements for baggage handling. This has an adverse effect on subsequent manual image interpretation and intelligent recognition. Summary of the Invention
[0004] Based on the above analysis, the present invention aims to provide a CT projection image enhancement method and apparatus to improve the spatial resolution of CT projection images.
[0005] The objective of this invention is mainly achieved through the following technical solutions: In a first aspect, embodiments of this application provide a CT projection image enhancement method, including: Acquire CT projection images of the target to be detected; The CT projection image is enhanced once using a first neural network model to eliminate the deformation and distortion of the CT projection image; A second neural network model is used to perform secondary image enhancement on the CT projection image after primary image enhancement, thereby improving the resolution of the CT projection image.
[0006] Furthermore, the step of performing image enhancement on the CT projection image using a first neural network model to eliminate distortion and warping of the CT projection image includes: The first neural network model takes the CT projection image as input and the repaired image as output; The restored image is an image captured by the CT detector with a fixed shooting direction, and the fixed shooting direction is the same as the projection direction of the CT projection image.
[0007] Furthermore, the training samples of the first neural network model include: CT projection images of multiple projection directions corresponding to the same detection object and repair images corresponding to each of the CT projection images.
[0008] Furthermore, during the training of the first neural network model, the method for obtaining the repaired image includes: Determine the initial placement of the test objects used during training; Determine the projection direction of the CT projection image; Based on the initial placement method, the shooting direction of the repaired image and / or the placement method of the detection object are determined according to the projection direction; Based on the shooting direction of the restored image and / or the placement method of the detection object, the restored image corresponding to the projection direction is obtained through the CT detector.
[0009] Furthermore, the step of performing secondary image enhancement on the CT projection image after primary image enhancement using a second neural network model to improve the resolution of the CT projection image includes: The second neural network model takes the CT projection image after the first image enhancement as input and outputs a high-resolution image; The resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is a detector with a resolution higher than that of a CT detector in the field of security inspection.
[0010] Further, acquiring the CT projection image of the target includes: Determine the projection direction of the CT projection image; The CT projection image is obtained according to the projection direction.
[0011] Further, based on the CT projection image, the equivalent metal atomic number of the metal region is determined, and the equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region in the CT projection image; A third neural network model is determined based on the equivalent metal atomic number. The third neural network model takes an image containing artifacts of the equivalent metal atomic number as input and an image not containing artifacts of the equivalent metal atomic number as output. The third neural network model is used to remove metal artifacts from the CT projection image.
[0012] Secondly, embodiments of this application provide a CT projection image enhancement device, including: an acquisition module, a first image enhancement module, and a second image enhancement module; The acquisition module is used to acquire CT projection images of the target being detected; The first image enhancement module is used to perform image enhancement on the CT projection image through a first neural network model to eliminate the deformation and distortion of the CT projection image; The second image enhancement module is used to perform secondary image enhancement on the CT projection image after primary image enhancement using a second neural network model, thereby improving the resolution of the CT projection image.
[0013] Furthermore, the first neural network model takes the CT projection image as input and the repaired image as output; The restored image is an image captured by the CT detector with a fixed shooting direction, and the fixed shooting direction is the same as the projection direction of the CT projection image.
[0014] Furthermore, the second neural network model takes the CT projection image after the first image enhancement as input and outputs a high-resolution image; The resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is an X-ray detector with a resolution higher than that of a CT detector in the art.
[0015] The technical solutions provided by the embodiments of the present invention have at least one of the following technical effects: 1. The first neural network model repairs the deformation and distortion of CT projection images, laying the foundation for obtaining high-quality CT projection images. The second neural network model improves the resolution of CT projection images, thereby improving the quality of CT projection images and enhancing the ability of CT detectors to identify objects. This enables the independent use of CT to detect items during security checks.
[0016] 2. By using a third network model, metal artifacts on the CT projection images after two enhancements are eliminated, further improving the quality of the CT projection images, so that CT can be used independently to detect metal objects and items containing metal during security checks.
[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description
[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0019] Figure 1 A flowchart of a CT projection image enhancement method provided in an embodiment of the present invention; Figure 2 This invention provides a schematic diagram of the orientation of each detection direction of the security inspection machine for embodiments of the invention; Figure 3 This is a schematic diagram of the structure of a CT projection image enhancement device provided in an embodiment of the present invention. Detailed Implementation
[0020] Embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.
[0021] Digital radiology (DR) imaging technology is highly mature and can clearly capture X-ray transmission images of the interior of objects at a specified angle. DR's most prominent advantages are its high resolution, producing clear, detailed images, and its wide application, making it a significant aid in current security checks. However, due to obstructions and limitations in the shooting angle, it is difficult to independently and effectively identify drugs, explosives, and biological tissues.
[0022] CT scans can accurately obtain the equivalent atomic number and electron density at each location in space through the reconstruction of high and low energy spectrum images, thus efficiently performing the task of substance identification. The advantage of CT projection images lies in their good ability to identify drugs, explosives, and biological tissues; however, the resolution of CT images is not high. Especially when applied to security inspection equipment, the resolution of CT projection images is even lower due to the influence of screw pitch and the security inspection time period, and the projected images may even be distorted.
[0023] Specifically, in security inspection scenarios, there are certain requirements for the bag-passing speed of CT scans, which necessitates the use of CT projection image acquisition technology with a larger pitch. This results in low resolution of the acquired CT sine wave image, and also leads to low resolution and some distortion in the CT reconstruction reprojection image, causing significant difficulties in image recognition and reprojection.
[0024] Furthermore, CT detectors are surround-type, requiring sampling from multiple angles within a short time. Therefore, they are designed as single-row detectors with smaller crystals, resulting in shorter light-sensing time and naturally lower image clarity. In contrast, DR detectors have fixed angles and larger crystals, allowing for more sufficient light-sensing time and thus clearer image quality. Consequently, the image clarity of detectors used in CT imaging differs significantly from that of DR detectors.
[0025] Based on the characteristics of CT and DR detectors, current technologies utilize both in security screening equipment. However, in real-world scenarios, DR detectors are fixed in position and can only capture images from a fixed angle. This means that if neither the DR nor the CT detector captures the true properties of the object being inspected from the same angle, manual inspection is the only option, significantly reducing security screening efficiency.
[0026] To address the aforementioned technical problems, embodiments of the present invention provide a CT projection image enhancement method, comprising the following steps: Step 1: Obtain the CT projection image of the target to be detected.
[0027] In existing technologies, when CT and DR detectors coexist, the DR detector is fixed in its configuration. Therefore, the projection direction of the slice image in existing technologies defaults to the imaging direction of the DR detector. However, in this embodiment, the security inspection equipment only has a CT detector, so it is necessary to first determine the projection direction of the CT projection image. Then, the CT projection image is acquired based on the projection direction. It should be noted that the projection direction of the CT projection image can be preset or temporarily specified. For example, to facilitate detection, the projection direction can be preset. When the CT projection image obtained in the preset direction cannot reflect the attributes of the target being detected, the inspector can manually set a new projection direction to observe the target from different directions.
[0028] Step 2: Using the first neural network model, perform image enhancement on the CT projection image to eliminate the deformation and distortion of the CT projection image.
[0029] Step 3: Using a second neural network model, perform secondary image enhancement on the CT projection image after the first image enhancement to improve the resolution of the CT projection image.
[0030] In this embodiment of the application, when using a CT detector alone to detect an object, based on the characteristics of the CT detector, the following two problems need to be solved: 1. Image distortion and warping issues caused by large-pitch rotational CT reconstruction and reprojection; 2. The CT detector has poor imaging quality.
[0031] In this embodiment, step 2 is used to solve problem 1, and step 3 is used to solve problem 2. Increasing the resolution cannot eliminate image distortion and warping, but eliminating image distortion and warping can partially improve the image resolution. Therefore, this application first eliminates image distortion and warping, and then increases the image resolution.
[0032] Specifically, regarding problem 1, the CT detector rotates at a certain pitch to obtain a CT sinusoidal image, which is the root cause of image distortion and warping in CT projection images. Therefore, in this embodiment, the first neural network takes the CT projection image as input and the repaired image as output to obtain the image captured by the CT detector with a fixed shooting direction, where the shooting direction is the same as the projection direction corresponding to the CT projection image. Through this method, image distortion and warping in the CT projection image are repaired as much as possible.
[0033] Different objects will exhibit varying degrees of deformation and distortion when projected in different directions. Some of these deformations and distortions are easy to repair, while others are not. If training is performed only from a single direction, it will lead to significant differences in the accuracy of different types of objects, ultimately reducing the overall computational accuracy of the model. For example, when the shooting direction is horizontal, the accuracy rates of the repaired images for objects A, B, and C are 60%, 80%, and 90%, respectively. If the above model is applied to a real-world scenario, object A will most likely not be detected.
[0034] To address the aforementioned issues, multiple CT projection images and corresponding repaired images were acquired for the same object under test, and all acquired projection images and repaired images were used as training samples.
[0035] Specifically, such as Figure 2 As shown, the direction through which the package passes is defined as the z-axis, the horizontal viewing angle as the x-axis, and the vertical viewing angle as the y-axis. In existing technologies, DR detectors capture images of the package from both horizontal and vertical viewing angles. Therefore, in this embodiment, the projection direction is assumed to be the horizontal and vertical viewing angles, and the direction of the image being repaired is also defined as the horizontal and vertical viewing angles.
[0036] Preferably, other projection angles can also be set based on the default projection direction, such as deviations of 30°-60° from the y-direction, deviations of 30°-60° from the x-direction, deviations of 30°-60° from the z-direction, and combinations of the above deviation angles.
[0037] For other projection angles, obtaining the corresponding repaired images requires adjusting the position of the CT detector in the security checkpoint, adjusting the placement of the object being inspected, or both, to ensure that the content displayed in the CT projection image is the same as what is observed when the CT detector is at a fixed angle. The specific adjustment process is as follows: Determine the initial placement of the detection objects used during training; determine the projection direction; based on the initial placement, determine the shooting direction of the restored image and / or the placement of the detection objects according to the projection direction.
[0038] After obtaining sufficient training samples, such as 10,000, in the manner described above, the first neural network model is then trained. In this embodiment, the first neural network model is a generative adversarial network model, comprising two parts: one part, called the generator, takes a CT projection image as input and a restored image as output; the other part, called the discriminator, takes the restored image and an undistorted image obtained by the generator as input and the score of the restored image as output. The score of the restored image characterizes the degree of deformation and distortion of the restored image. For example, for the same object, image A is an undistorted image, and image B is a restored image. The discriminator determines whether image B is deformed by comparing the image features of the two. The more deformed or distorted features, the lower the score, and vice versa.
[0039] During training, the discriminator is first trained to distinguish between repaired images and images without deformation or distortion based on rule A. The discriminator's parameters are then fixed so that it can only make judgments based on rule A. Next, the generator is trained to produce repaired images that satisfy rule A. When the discriminator's output score reaches a preset value, it indicates that the repaired image satisfies rule A. At this point, the generator's parameters are fixed so that it can only generate repaired images based on rule A. The discriminator is then trained again to distinguish between repaired images and images without deformation or distortion based on both rule A and rule B. The discriminator's parameters are again fixed so that it can only make judgments based on rule A and rule B, and so on. This process is iterated until the discriminator can no longer generate new rules.
[0040] Specifically, 10,000 reconstructed and reprojected images Xi (the default projection image represents the initial state of the restored image) and 10,000 corresponding fixed-angle CT images were collected as sample data Yi (images without deformation or distortion): a. First, train the discriminator D1: Input the data Y1, Y2...Y10000 and X1, X2...X10000 into D1, and adjust D1 to maximize the total score V using gradient descent. maxD1 The general formula for the total score of D1 is: , where m represents the total number of samples.
[0041] b. Retrain the generator G1: Input the data X1, X2...X10000 into G1 to obtain X1', X2'...X10000'.
[0042] c. Keeping the parameters of D1 unchanged, the discriminator D1 takes X1', X2'...X10000' and data Y1, Y2...Y10000 as inputs and the total score of G1 as output. The general formula for the total score of G1 is: .
[0043] d. Adjust the maximum total score Vmax of G1 using gradient descent. G1 So that V G The preset value has been reached.
[0044] e. Keeping the parameters of G1 unchanged, train D2 with X1', X2'...X10000' and data Y1, Y2...Y10000 as input, and adjust D2 to maximize the total score V using gradient descent. maxD2 The general formula for the total score of D2 is: .
[0045] f. Repeat execution of ae to obtain the Nth generation generators GN and DN, when V DN and V D(N-1) When the difference does not exceed the preset range, the current process ends.
[0046] During use, data is input into GN, the corrected image is obtained through GN, and the image obtained by GN is used in the next step of calculation.
[0047] In the embodiments of this application, when the first neural network model is actually used, if multiple CT projection images are input for the same detection target, the image with the best enhancement effect is selected as the input of the second neural network model based on the image enhancement results of each projection direction of each CT projection image.
[0048] In this embodiment, after eliminating image distortion and deformation, the second neural network model takes the enhanced CT projection image as input and outputs a high-resolution image. The resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is a detector with a higher resolution than the CT detector in the security inspection field. That is, when improving the resolution of the CT projection image, this application not only uses the resolution of the DR detector as a reference, but also other detectors in the field, such as terahertz detectors.
[0049] It should be noted that during training, the input to the second neural network is the image captured by the CT detector with a fixed shooting direction, and the output of the second neural network is the high-resolution image captured by the high-resolution detector with the same shooting direction. Since the second neural network model is mainly used to improve resolution and does not involve changing the image shape, it is not necessary to set multiple projection angles when training the second neural network model. It is only necessary to ensure that the shooting angles of the input CT projection image and the output high-resolution image are the same.
[0050] The second neural network model is trained with a convolutional autoencoder to obtain the autoencoder.
[0051] An autoencoder consists of an encoder and a decoder, both of which are trained using a convolutional neural network algorithm.
[0052] The encoder consists of an input layer, convolutional layers, pooling layers, and an output layer: The input layer contains CT projection images F1(a11,a12,a13...)...FN(a11,a12,a13...) acquired from multiple projection directions. The convolutional layer consists of four convolutional layers with a kernel size of 3×3 and a stride of 2, and is used to extract features from the input data from the input layer. The pooling layer uses average pooling with a step size of 2; The output layer is the result T1(w1,w2,w3...)...TN(w1,w2,w3...) after extracting features from the convolutional and pooling layers, and provides input for the next step of the decoder; The decoder consists of an input layer, an unpooling layer, a deconvolution layer, and an output layer: The input layer receives the output data from the encoder described above. The anti-pooling layer is an anti-average pooling layer with a step size of 2; The deconvolutional layer consists of 4 layers of deconvolutional layers with a kernel size of 3×3 and a stride of 2. The output layer is the result G'1(a11,a12,a13...)...G'N(a11,a12,a13...) after restoring the features of the depooling layer and the deconvolution layer; Loss function used in training:
[0053] Where G'i is the image vector reconstructed after the encoder de-encoder, and Gi is the repaired image corresponding to the CT projection image.
[0054] When the target object is metal, metal artifacts will appear in the CT projection image. To ensure image processing quality, it is necessary to remove these metal artifacts from the CT projection image. The following problems exist in the process of removing metal artifacts: First, different metals produce artifacts of varying sizes. For image processing, the difficulty of accurately determining artifact boundaries varies depending on the artifact size; generally, the larger the artifact area, the easier its boundary is to define. Second, when the metal region of the target object attenuates X-rays more than other regions, metal artifacts appear at the boundary between the metal region and other areas. Different metals attenuate X-rays to varying degrees, resulting in slightly different image characteristics of the artifacts, such as artifact brightness and the smoothness of artifact boundaries. These factors all have varying degrees of impact on subsequent image processing results. Finally, the materials of metal parts in bags and knives are usually alloys or treated with gold plating. Therefore, in the above cases, metal artifacts are determined by multiple elements, requiring a chemical analysis to distinguish the influence of different substances on metal artifacts.
[0055] To address the aforementioned technical problems, following step 3, this application provides a method for removing metal artifacts from CT projection images using a neural network model, including: Step S1: Determine the equivalent metal atomic number of the metal region based on the CT projection image.
[0056] In this embodiment, the metal equivalent atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT projection image. Specifically, the atomic number of each pixel in the CT projection image is calculated, and the metal region in the CT projection image is determined based on the atomic number. Then, the average atomic number of each pixel in each metal region is determined as the equivalent metal atomic number.
[0057] The equivalent metallic atomic number is a fixed value or falls within a fixed range of variation. Specifically, in the field of alloys, the components of an alloy typically vary within a certain range. For example, a high-strength ferritic steel might have the following partial chemical composition: Ni 2–15%, Mn 2–10%, Al 1–6%, Cu 1.5–4%, and Cr 8–12%. In this case, the equivalent metallic atomic number is within a fixed range. Conversely, if the partial chemical composition is: Ni 2%, Mn 10%, Al 6%, Cu 4%, and Cr 12%, the equivalent metallic atomic number is also a fixed value.
[0058] Step S2: Determine the third neural network model based on the equivalent metal atomic number.
[0059] In this embodiment, the third neural network model takes an image containing artifacts of equivalent metal atomic numbers as input and an image without artifacts of equivalent metal atomic numbers as output. Based on the equivalent metal atomic numbers, the third neural network model is typically categorized into three types: copper-iron-nickel-chromium and alloys, titanium and titanium alloys, and platinum.
[0060] In this embodiment, step 3 obtains a high-resolution CT projection image, which facilitates more accurate determination of artifact boundaries and equivalent metal atomic numbers during metal artifact removal, thereby improving the metal artifact removal effect. The third neural network model can be a convolutional autoencoder or a VAE algorithm.
[0061] The training method for the third neural network model is as follows: Based on a preset model atomic number or a model atomic number range, obtain an artifact-free CT slice image. The CT slice image includes a metal region corresponding to at least one preset model atomic number or at least one preset model atomic number range. Metal artifacts are added to CT slice images based on the preset atomic number of the model. CT slice images without artifacts and CT slice images with added metal artifacts were used as training samples for the corresponding neural network models.
[0062] This application provides a CT projection image enhancement device based on DR images, including: an acquisition module 301, a first image enhancement module 302, and a second image enhancement module 303; The acquisition module 301 is used to acquire the CT projection image of the target being detected; The first image enhancement module 302 is used to perform image enhancement on the CT projection image through the first neural network model to eliminate the deformation and distortion of the CT projection image; The second image enhancement module 303 is used to perform secondary image enhancement on the CT projection image after primary image enhancement through a second neural network model, thereby improving the resolution of the CT projection image.
[0063] In this embodiment of the application, the first neural network model takes a CT projection image as input and a repaired image as output; the repaired image is an image captured by a CT detector when the shooting direction is fixed, and the fixed shooting direction is the same as the projection direction of the CT projection image.
[0064] In this embodiment, the second neural network model takes a CT projection image after one image enhancement as input and a high-resolution image as output; the resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is an X-ray detector with a resolution higher than that of a CT detector in the art.
[0065] To illustrate the feasibility of the above solution, the following embodiment is provided: Example 1 This application embodiment uses the aforementioned training method to obtain a first neural network model and a second neural network model. Then, data from 1000 packages are collected, and four sets of experiments are conducted: Group 1: Identify 1000 packages using DR images and commonly used dangerous goods identification models.
[0066] In the embodiments of this application, the hazardous materials identification models that can be used include: YOLOV5, RCNN, or SSD.
[0067] The second group: Identify 1,000 packages using unprocessed CT projection images and commonly used hazardous materials identification models; The third group: using unprocessed CT projection images and the first neural network model of this application to identify 1000 packages; Group 4: Using unprocessed CT projection images, the first neural network model and the second neural network model of this application, 1000 packages were identified.
[0068] This application uses the positive detection rate to evaluate the recognition results of the four groups, as shown in Table 1: Table 1. Positive detection rate for each group
[0069] As shown in Table 1, the first neural network model can indeed eliminate image distortion and warping, thereby increasing the positive detection rate by nearly 5 times. However, due to the low resolution, the positive detection rate of the third group is still lower than that of the first and fourth groups. At the same time, the positive detection rate using the first neural network model and the second neural network model (the technical solution of this application) is increased to 97%, which is basically the same as the positive detection rate of the first group, indicating that the technical solution of this application can solve problems 1 and 2, thereby realizing the independent use of CT detectors for security checks.
[0070] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for enhancing CT projection images, characterized in that, include: Acquire a CT projection image of the target to be detected; wherein the CT projection image is obtained by large-pitch rotational CT reconstruction and reprojection; The CT projection image is enhanced once using a first neural network model to eliminate the deformation and distortion of the CT projection image; wherein, the first neural network model includes a generator and a discriminator; A second neural network model is used to perform secondary image enhancement on the CT projection image after primary image enhancement, so as to improve the resolution of the CT projection image. The step of performing image enhancement on the CT projection image using a first neural network model to eliminate distortion and warping of the CT projection image includes: The first neural network model takes the CT projection image as input and the repaired image as output; The restored image is an image captured by the CT detector with a fixed shooting direction, and the fixed shooting direction is the same as the projection direction of the CT projection image; The step of performing secondary image enhancement on the CT projection image after primary image enhancement using a second neural network model to improve the resolution of the CT projection image includes: The second neural network model takes the CT projection image after the first image enhancement as input and outputs a high-resolution image; The resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is a detector with a higher resolution than a CT detector in the field of security inspection. The training samples of the first neural network model include: CT projection images of the same object in multiple projection directions and repaired images corresponding to each of the CT projection images; During the training of the first neural network model, the method for obtaining the repaired image includes: Determine the initial placement of the test objects used during training; Determine the projection direction of the CT projection image; Based on the initial placement method, the shooting direction of the repaired image and / or the placement method of the detection object are determined according to the projection direction; Based on the shooting direction of the restored image and / or the placement method of the detection object, the restored image corresponding to the projection direction is obtained through the CT detector.
2. The method according to claim 1, characterized in that, The acquisition of the CT projection image of the target to be detected includes: Determine the projection direction of the CT projection image; The CT projection image is obtained according to the projection direction.
3. The method according to any one of claims 1 to 2, characterized in that, The detection target is metal, and the method further includes: Based on the CT projection image, the equivalent metal atomic number of the metal region is determined. The equivalent metal atomic number is used to characterize the average atomic number of each pixel in the metal region of the CT projection image. A third neural network model is determined based on the equivalent metal atomic number. The third neural network model takes an image containing artifacts of the equivalent metal atomic number as input and an image not containing artifacts of the equivalent metal atomic number as output. The third neural network model is used to remove metal artifacts from the CT projection image.
4. A CT projection image enhancement device, characterized in that, include: The module includes an acquisition module, a first image enhancement module, and a second image enhancement module. The acquisition module is used to acquire CT projection images of the target being detected; wherein, the CT projection images are obtained by large-pitch rotational CT reconstruction and reprojection; The first image enhancement module is used to perform image enhancement on the CT projection image through a first neural network model to eliminate the deformation and distortion of the CT projection image; wherein, the first neural network model includes a generator and a discriminator; The second image enhancement module is used to perform secondary image enhancement on the CT projection image after primary image enhancement using a second neural network model, thereby improving the resolution of the CT projection image; The step of performing image enhancement on the CT projection image using a first neural network model to eliminate distortion and warping of the CT projection image includes: The first neural network model takes the CT projection image as input and the repaired image as output; The restored image is an image captured by the CT detector with a fixed shooting direction, and the fixed shooting direction is the same as the projection direction of the CT projection image; The step of performing secondary image enhancement on the CT projection image after primary image enhancement using a second neural network model to improve the resolution of the CT projection image includes: The second neural network model takes the CT projection image after the first image enhancement as input and outputs a high-resolution image; The resolution of the high-resolution image is based on the resolution of the high-resolution detector, which is a detector with a higher resolution than a CT detector in the field of security inspection. The training samples of the first neural network model include: CT projection images of the same object in multiple projection directions and repaired images corresponding to each of the CT projection images; During the training of the first neural network model, the method for obtaining the repaired image includes: Determine the initial placement of the test objects used during training; Determine the projection direction of the CT projection image; Based on the initial placement method, the shooting direction of the repaired image and / or the placement method of the detection object are determined according to the projection direction; Based on the shooting direction of the restored image and / or the placement method of the detection object, the restored image corresponding to the projection direction is obtained through the CT detector.
Citation Information
Patent Citations
A method of removing metal artifacts from CT images based on image inpainting
CN109472754A
Training method of super-resolution reconstruction network model and scanning image processing method
CN113361689A
