A three-dimensional rapid reconstruction system for dual-view X-ray fluoroscopic imaging

Abstract

The present invention relates to the field of engine testing, and specifically to a three-dimensional rapid reconstruction system for dual-view X-ray fluoroscopic imaging. It includes: a dual-view X-ray imaging system and dual-view data acquisition, using orthogonal ray arrangement to carry out X-ray data acquisition; based on dual-view projection three-dimensional structural feature extraction and cross-dimensional fusion, the representation generation network model is used as the basic framework, the dual-view projection image is subjected to feature extraction through a residual network, the number of channels of the feature map is gradually increased by using cascaded convolution operations, and the two-dimensional projection domain features are mapped to three-dimensional space domain features; feature-based three-dimensional rapid reconstruction optimization, based on the two-dimensional feature values ​​of the dual-view projection data, by encoding the three-dimensional depth information and the channel information in the same dimension, and using two-dimensional convolution instead of three-dimensional convolution features for processing and reconstruction, the present invention improves the spatial resolution of reconstruction to realize dynamic profile testing of the engine combustion process.

Description

Technical Field

[0001] The invention relates to the field of engine testing, and in particular to a three-dimensional rapid reconstruction system for dual-viewing angle X-ray fluoroscopic imaging. Background Art

[0002] The CT imaging mode of non-destructive testing can reconstruct a two-dimensional tomographic image or a three-dimensional stereoscopic image of an object without damaging the object by using the attenuation information of X-rays penetrating the object. It can intuitively characterize the internal structure of the object and has broad application prospects in aerospace, precision material manufacturing and other fields, especially in the structural identification of objects.

[0003] The dynamic single-view projection test technology during the engine test process is used to dynamically test the retreat state of the combustion surface of the end-burning solid rocket engine. During the ground ignition test of the end-burning solid rocket engine, a DR device with a viewing angle is arranged horizontally to test the dynamic combustion surface parameters of the engine. The dynamic single-view projection test technology during the engine test process only contains the object information in the projection direction. Under the condition of unknown reconstructed objects, the extremely sparse projection image contains insufficient information to describe the spatial structure of the object. Therefore, the DR test method is not applicable to core-hole surface combustion engines.

[0004] In specific environments, such as dynamic testing of structural profile changes within solid rocket engine tests, traditional CT reconstruction is limited by the number of projections required, and the huge amount of computation also limits the reconstruction speed. Considering that objects cannot be rotated in the test environment, multi-view ray-detectors cannot be arranged to acquire complete circumferential projection data, and it is difficult to quickly obtain projection data sufficient for reconstruction in a short period of time, it is impossible to quickly monitor and identify internal structural changes. Summary of the invention

[0005] Technical issues to be solved

[0006] The present invention provides a three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging to solve the problem that the existing solid rocket engine ground test DR test method cannot meet the test requirements of the engine internal combustion profile.

[0007] To solve the technical problem, the present invention adopts the technical solution

[0008] A three-dimensional rapid reconstruction system for dual-view X-ray fluoroscopic imaging, comprising:

[0009] Dual-view X-ray imaging system and dual-view data acquisition. The dual-view X-ray imaging system adopts an orthogonal ray arrangement to ensure that the ray source and the detector can work synchronously with consistent parameters, synchronously capture the instantaneous state of the section, determine the component indicators related to the X-ray data acquisition system, the geometric parameters and physical parameters of the imaging system, and carry out the hardware construction technology of the X-ray data acquisition system;

[0010] Based on the extraction and cross-dimensional fusion of three-dimensional structural features of dual-view projection, the representation generation network model is used as the basic framework. The residual network is used to extract features from the dual-view projection image to accelerate network training and avoid gradient disappearance. At the same time, the number of channels of the feature map is gradually increased by cascaded convolution operations, and the two-dimensional projection domain features are mapped to three-dimensional spatial domain features. The dual-view spatial domain features are fused by concatenating features and then convolving them.

[0011] Feature-based three-dimensional fast reconstruction optimization is based on the two-dimensional eigenvalues ​​of dual-view projection data. By encoding the three-dimensional depth information and channel information in the same dimension, two-dimensional convolution is used instead of three-dimensional convolution features for processing and reconstruction, thereby improving the spatial resolution of reconstruction and realizing dynamic profile testing of the engine combustion process.

[0012] Furthermore, the three-dimensional data feature extraction and reconstruction adopts three-dimensional convolution and deconvolution as basic operations.

[0013] Furthermore, the representation network in the representation generation network model is composed of residual blocks and maximum pooling. Three residual blocks are combined with maximum pooling to construct a downsampling block, and each downsampling block reduces the spatial size of the feature map by 2 times.

[0014] Furthermore, the generative network in the representation generation network model improves the spatial resolution of the feature map through two-dimensional deconvolution, and combines the deconvolution with three residual blocks to construct upsampling blocks, each of which increases the spatial size of the feature map by 2 times.

[0015] Furthermore, the representation network and the generation network in the representation generation network model use convolution operations of the same dimension, which facilitates feature sharing between different network structures.

[0016] Furthermore, in the representation generation network model, the number of two-dimensional convolution kernel channels is controlled to be twice the data space resolution, and finally the three-dimensional feature space size is reduced through the convolution operation to ensure the sparsity of the features, thereby avoiding the loss of feature information in the maximum pooling operation.

[0017] Furthermore, the representation network in the representation network model further abstracts and encodes the spatial features. During the decoding process, the generation network gradually reconstructs the spatial features into three-dimensional density data. Through the training process, the generated three-dimensional data is approximated to the real data to constrain the abstraction process of the representation network, thereby ensuring the accuracy and completeness of the extracted features.

[0018] Beneficial effects obtained by the present invention

[0019] The present invention can realize the three-dimensional rapid reconstruction of the cross-sectional structure of the engine during dynamic operation based on the two-dimensional data of the engine dynamic projection from two perspectives, realize the dynamic cross-sectional test of the engine combustion process; realize the dynamic test of the core hole charge solid rocket engine test combustion cross-sectional data during the engine test. The present invention adopts a deep learning model in the reconstruction process, which is to learn and train the implicit relationship in the data, implicitly obtain prior knowledge in the training process, and there is no need to establish a complex mathematical model for the data distribution, which greatly simplifies the use of prior knowledge. For the feature extraction, fusion and cross-dimensional mapping of dual-perspective projection data, the design adopts the "representation generation network model" as the basic framework, and extracts features from the dual-perspective projection image through the residual network, accelerates network training, avoids gradient disappearance, and uses cascaded convolution operations to gradually increase the number of channels of the feature map, maps the two-dimensional projection domain features to three-dimensional space domain features, and fuses the dual-perspective space domain features by concatenating features and then convolving. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0021] Figure 1 : A flow chart of a three-dimensional rapid reconstruction method for dual-view X-ray fluoroscopic imaging;

[0022] Figure 2 : Schematic diagram of the imaging system;

[0023] Figure 3 : Schematic diagram of dual-view reconstruction network structure;

[0024] Figure 4 : Residual learning;

[0025] Figure 5 : Comparison chart of 3D convolution and 2D convolution;

[0026] Figure 6 : Schematic diagram of downsampling block (a) and upsampling block (b). DETAILED DESCRIPTION

[0027] The technical problem to be solved by the present invention is that the existing DR test method for ground tests of solid rocket engines cannot meet the test requirements of the internal combustion profile of the engine, and a new three-dimensional rapid reconstruction method based on data-driven dual-view X-ray fluoroscopic imaging is provided.

[0028] This method consists of the following three modules:

[0029] Design of fast perspective projection data acquisition system and dual-perspective data acquisition: This module starts with the hardware. Based on the research on the X-ray test method of the internal structure profile of the engine, it determines the component indicators related to the X-ray data acquisition system, the geometric parameters and physical parameters of the imaging system, and conducts research on the hardware construction technology of the X-ray data acquisition system.

[0030] Based on dual-view projection three-dimensional structure feature extraction and cross-dimensional fusion: The dual-view projection three-dimensional structure feature extraction and cross-dimensional fusion refers to the feature extraction of dual-view projection images through a residual network, accelerating network training and avoiding gradient disappearance, while using cascaded convolution operations to gradually increase the number of channels of the feature map, mapping the two-dimensional projection domain features into three-dimensional space domain features, and fusion of the dual-view space domain features by feature splicing and then convolution.

[0031] Feature-based 3D fast reconstruction optimization technology:

[0032] The feature-based three-dimensional fast reconstruction optimization technology refers to a two-dimensional feature value based on dual-view projection data. By encoding the three-dimensional depth information and channel information in the same dimension (channel dimension), two-dimensional convolution is used instead of three-dimensional convolution features for processing and reconstruction, which greatly improves the computational efficiency and improves the spatial resolution of reconstruction to realize dynamic profile testing of the engine combustion process.

[0033] In order to make the purpose, features and advantages of the technical solution proposed in the present invention more obvious and easy to understand, the embodiments of the technical solution proposed in the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the proposed technical solution, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0034] The use of a vertical dual-viewing method to synchronously acquire dynamic DR data of the engine's transverse profile is the first time this has been tested in an engine test.

[0035] like Figure 1 The technical route of the present invention is shown, which includes offline 3D reconstruction network training, online real-time data acquisition and online real-time application of 3D reconstruction network, which is divided into the following three key technologies:

[0036] Design of dual-view projection data fast acquisition system and dual-view data acquisition

[0037] The dual-view X-ray imaging system adopts an orthogonal ray arrangement scheme (such as Figure 2 As shown in the figure, the ray paths are arranged orthogonally, thereby ensuring the orthogonality of the object feature distribution in the dual-view projection data. On this basis, in order to ensure the consistency of dual-view imaging, a study on the timing control of dual-view synchronous imaging is carried out. Through hardware design, it is ensured that the ray source and the detector can work synchronously with consistent parameters and synchronously capture the instantaneous state of the profile.

[0038] In order to ensure the big data support problem in the reconstruction network training process, complete CT projection data is obtained before the object is equipped, and the reconstruction network is trained using it during the training phase. The use of relatively complete data can improve the generalization ability of the network to reconstruct different forms. In the test phase, the actual collected dual-view projection data is used for reconstruction, and the three-dimensional object is reconstructed using the actual data.

[0039] Three-dimensional structural feature extraction and cross-dimensional fusion based on dual-view projection

[0040] The projection data under dual-view is the integral signal of the structure in different directions of the object. The key to network reconstruction depends on the accuracy of feature extraction and dual-view feature matching fusion. For the feature extraction, fusion and cross-dimensional mapping of dual-view projection data, the design adopts the "representation generation network model" as the basic framework, such as Figure 3 shown.

[0041] The residual network is used to extract features from the dual-view projection images, accelerate network training, and avoid gradient disappearance. At the same time, the number of channels of the feature map is gradually increased by using cascaded convolution operations, and the two-dimensional projection domain features are mapped to three-dimensional spatial domain features. The dual-view spatial domain features are fused by concatenating and then convolving the features. The representation network further abstracts and encodes the spatial features, and the generation network gradually reconstructs the spatial features into three-dimensional density data during the decoding process. The generated three-dimensional data is approximated to the real data through the training process to constrain the abstraction process of the representation network, thereby ensuring the accuracy and completeness of the extracted features.

[0042] In the representation generation network framework, the representation network gradually compresses the feature space size and extracts the main components of the feature. This process often leads to the expansion of the network receptive field and the loss of detail information, which is manifested as incomplete reconstruction of small structures during the reconstruction process. Therefore, on the basis of feature extraction, multi-scale feature fusion is performed, and cross-layer connections are used to transfer features between network structures of the same scale (such as Figure 3At the same time, in order to improve the network's expressiveness, the feature processing at the cross-layer connection adopts a solution of concatenating the features and then fusing them using convolution. The cross-layer connection introduces features of different scales into the reconstruction process, allowing the network to locate features at different resolutions, thereby enhancing the network's ability to reconstruct details.

[0043] Represents a network consisting of residual blocks ( Figure 4 ) and maximum pooling. The three residual blocks are combined with the maximum pooling to construct a downsampling block. Each downsampling block reduces the spatial size of the feature map by 2 times, as shown in Figure 6 (a) shows that the generative network improves the spatial resolution of the feature map through two-dimensional deconvolution. The deconvolution is combined with three residual blocks to construct an upsampling block, as shown in Figure 6 As shown in (b), each upsampling block increases the spatial size of the feature map by a factor of 2. Using convolution operations of the same dimension in the representation network and the generation network facilitates feature sharing between different network structures. This allows us to directly perform cross-layer connections between different network blocks with the same spatial resolution without having to crop or transform the features. In the representation generation network, by controlling the number of two-dimensional convolution kernel channels to twice the spatial resolution of the data, the convolution operation is finally used to reduce the size of the three-dimensional feature space to ensure the sparsity of the features, thereby avoiding the loss of feature information in the maximum pooling operation.

[0044] Feature-based 3D fast reconstruction optimization

[0045] On the basis of completing the design of the basic network model, we will conduct research on network operation optimization and training efficiency optimization to improve the network reconstruction efficiency and training speed. For three-dimensional data feature extraction and reconstruction, three-dimensional convolution and deconvolution are mostly used as basic operations, such as Figure 5 As shown in (a), the feature data and filter size processed in the three-dimensional convolution are both four-dimensional, which makes the generated network have a huge amount of parameters to be trained and the amount of calculation. Under the existing hardware resources, the three-dimensional convolution has a huge calculation and time overhead, which limits the spatial resolution and reconstruction speed of the reconstruction. At the same time, network optimization research was conducted on the number of network parameters and network structure. By encoding the three-dimensional depth information and channel information in the same dimension (channel dimension), as shown in Figure 5 As shown in (b), two-dimensional convolution is used instead of three-dimensional convolution features for processing and reconstruction, which greatly improves the computational efficiency and increases the spatial resolution of reconstruction.

[0046] The use of two-dimensional convolution can greatly reduce the number of parameters to be trained and the amount of computation required for convolution operations, thereby shortening the network's computing speed and the training time of the network.

[0047] In order to improve the expressive power of the network, the features are not directly added in the cross-layer connection. Instead, the features are concatenated and then fused using convolution.

Claims

1. A three-dimensional rapid reconstruction system for dual-view X-ray fluoroscopic imaging, characterized in that: include: Dual-view X-ray imaging system and dual-view data acquisition. The dual-view X-ray imaging system adopts an orthogonal ray arrangement to ensure that the ray source and the detector can work synchronously with consistent parameters, synchronously capture the instantaneous state of the section, determine the component indicators related to the X-ray data acquisition system, the geometric parameters and physical parameters of the imaging system, and carry out research on the hardware construction technology of the X-ray data acquisition system. Before the object is equipped, complete CT projection data is obtained, and the reconstruction network is trained using it in the training phase. In the test phase, the actual collected dual-view projection data is used for reconstruction, and the three-dimensional object is reconstructed using the actual data; Based on the extraction and cross-dimensional fusion of three-dimensional structural features of dual-view projection, the representation generation network model is used as the basic framework. The residual network is used to extract features from the dual-view projection image to accelerate network training and avoid gradient disappearance. At the same time, the number of channels of the feature map is gradually increased by cascaded convolution operations, and the two-dimensional projection domain features are mapped to three-dimensional spatial domain features. The dual-view spatial domain features are fused by concatenating features and then convolving them. Feature-based three-dimensional fast reconstruction optimization is based on the two-dimensional eigenvalues ​​of dual-view projection data. By encoding the three-dimensional depth information and channel information in the same dimension, two-dimensional convolution is used instead of three-dimensional convolution for processing and reconstruction, and the spatial resolution of reconstruction is improved to realize dynamic profile testing of the engine combustion process.

2. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to claim 1, characterized in that: The dual-view projection three-dimensional structure feature extraction and cross-dimensional fusion as well as the feature-based three-dimensional fast reconstruction optimization adopt two-dimensional convolution and two-dimensional deconvolution as basic operations.

3. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to claim 1, characterized in that: The representation network in the representation generation network model is composed of residual blocks and maximum pooling. Three residual blocks are combined with maximum pooling to construct downsampling blocks, and each downsampling block reduces the spatial size of the feature map by 2 times.

4. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to claim 1, characterized in that: The generative network in the representation generation network model improves the spatial resolution of the feature map through two-dimensional deconvolution, and combines the two-dimensional deconvolution with three residual blocks to construct upsampling blocks, each of which increases the spatial size of the feature map by 2 times.

5. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to any one of claims 1 to 4, characterized in that: The representation network and the generation network in the representation generation network model use convolution operations of the same dimension, which facilitates feature sharing between different network structures.

6. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to claim 1, characterized in that: In the representation generation network model, the number of two-dimensional convolution kernel channels is controlled to be twice the data space resolution, and finally the three-dimensional feature space size is reduced through the convolution operation to ensure the sparsity of the features, thereby avoiding the loss of feature information in the maximum pooling operation.

7. The three-dimensional rapid reconstruction system of dual-view X-ray fluoroscopic imaging according to claim 1, characterized in that: The representation network in the representation generation network model further abstracts and encodes the spatial features. During the decoding process, the generation network gradually reconstructs the spatial features into three-dimensional density data. Through the training process, the generated three-dimensional data is approximated to the real data to constrain the abstraction process of the representation network, thereby ensuring the accuracy and completeness of the extracted features.

Images