A sparse angle THz-CT image reconstruction method based on deep learning
Through the sparse angle THz-CT image reconstruction method based on deep learning, the improved UNet neural network and channel attention mechanism are used to solve the problem of reconstructing the THz-CT image under sparse angle sampling, and the high-resolution three-dimensional Reconstruct effect.
Patent Information
- Application Number
- CN202311486873.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-09
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2043-11-09
AI Technical Summary
The existing THz-CT imaging system is prone to blur and distortion when image reconstruction under sparse angle sampling, and the data usage efficiency is low, making it difficult to achieve high-resolution three-dimensional reconstruction.
The sparse angle THz-CT image reconstruction method based on deep learning is adopted, and the improved UNet neural network model is used to combine enhanced convolutional blocks, channel attention mechanism and regularization loss function to eliminate artifact processing and reconstruction of sparse angle images.
The reconstruction quality of sparse angle THz-CT images is significantly improved, the clarity of image edge profiles and features is enhanced, spatial resolution is improved, and high-quality reconstruction images can be restored at extremely sparse angles.
Smart Images

Figure CN117523095B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of terahertz nondestructive testing and image processing, and in particular to a sparse angle THz-CT image reconstruction method based on deep learning. Background Art
[0002] Three-dimensional terahertz computed tomography (THz-CT) imaging technology can visualize the internal information of an object and is an effective means of non-destructive testing of non-polar materials. However, for terahertz imaging systems, imaging systems that combine high-power transmitters with detector arrays are still under further study. Most of the current THz-CT imaging systems acquire data in a point scanning acquisition method, that is, THz-CT images acquired in a dense angle sampling method are directly reconstructed in three dimensions. It usually takes more than 10 hours to fully scan an object of several cubic centimeters. Direct three-dimensional reconstruction of THz-CT images acquired in a sparse angle sampling method often leads to blurring and distortion of the tomographic image, resulting in low data utilization efficiency of the sparse angle THz-CT reconstruction method. Therefore, it is necessary to explore a method that can use sparse angle THz-CT data for three-dimensional reconstruction.
[0003] To this end, many researchers have begun to explore reconstruction methods for sparse projection data. A large number of studies have drawn on mature algorithms and improved techniques in X-CT imaging to improve the resolution and accuracy of THz-CT reconstruction results. For example, the use of simultaneous algebraic reconstruction technology and ordered subset expectation maximization method to replace the traditional filtered back-projection method can effectively alleviate beam hardening; the total variation iteration method (TV) with gradient smoothing regularization term and the dictionary denoising algorithm based on K-SVD training can obtain higher quality reconstruction results. Recently, the widespread application of deep learning methods has provided a cutting-edge reconstruction idea for THz-CT reconstruction, which has obvious advantages in solving the underdetermined problem of sparse reconstruction. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-resolution THz-CT method for extremely sparse angles. The THz-CT image reconstruction algorithm can realize the reconstruction of extremely sparse angle THz-CT images, solving the problems of image blur and distortion under sparse reconstruction.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] The present invention relates to a sparse angle THz-CT image reconstruction method based on deep learning, which specifically includes the following steps:
[0007] (S1) collecting sparse angle THz-CT images and dense angle THz-CT images of training samples, wherein the sparse angle THz-CT images of the training samples in the thickness direction constitute an image set, and the dense angle THz-CT images of the training samples in the thickness direction constitute a label set, and the training set includes a label set and an image set;
[0008] (S2) inputting the training set into the neural network model for training to obtain a trained neural network model;
[0009] The neural network is an improved UNet network with a neural network structure depth of 4 layers, including enhanced convolution blocks (Enhanced Conv), basic convolution blocks (Basic Conv), pooling layers (Pooling), and upsampling layers (Upsampling). The encoder part consists of an enhanced convolution block and a pooling layer to form a downsampling module. After multiple operations, abstract features are obtained. Then the decoder part consists of an upsampling layer and a basic convolution block to form an upsampling module. After multiple operations, the target image is restored from the abstract features. Finally, a layer of 1*1 convolution is performed to reduce the dimension to 1 channel to obtain the target output. The convolution layers of corresponding dimensions in the encoder and decoder are channel-spliced. The enhanced convolution block (Enhanced Conv) of the encoder consists of a 3*3 convolution block, an asymmetric convolution (ACM Conv), a channel attention mechanism (SE Block), a linear rectification activation function, and a batch normalization layer. The basic convolution block (Basic Conv) consists of two layers of 3*3 convolution blocks, a linear rectification activation function and a batch normalization layer; the introduced channel attention mechanism is divided into two parts: Extract global feature extraction and Transform channel weighted manifold transformation, which are introduced in the 3*3 convolution block and the asymmetric convolution block, and the adjacent channel attention features are interconnected; the number of downsampling channels of the overall network is 64, 128, 256, and 512 respectively, and the upsampling is the opposite;
[0010] (S3) inputting the sparse angle THz-CT image to be reconstructed into the trained neural network model to obtain a high-resolution THz-CT reconstructed image;
[0011] (S4) The THZ-CT reconstructed image corresponding to the sparse angle THz-CT image of the entire height of the sample to be reconstructed is deeply stacked along the plane axis of the THZ-CT reconstructed image to obtain a three-dimensional image of the sample to be reconstructed.
[0012] Specifically, the sparse angle THz-CT image acquisition method of step (S1) and step (S3) is as follows: when the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample section is collected by sparse angle sampling to obtain a sinusoidal image, the sparse terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the sinusoidal image is processed by a filtered back projection algorithm to obtain a slice image of the corresponding section, that is, a sparse angle THz-CT image.
[0013] Similarly, the dense-angle THz-CT image acquisition method in step (S1) is as follows: every time the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample cross section is collected through dense-angle sampling, the dense terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the preprocessed sinusoidal image is processed by a fast shrinkage threshold soft iterative algorithm to obtain a dense-angle THz-CT image.
[0014] Specifically, the preprocessing includes filtering, truncation, multiple averaging, background denoising and spectrum power integration.
[0015] Specifically, in step (S2), the neural network uses the dense angle THz-CT image as a label image and the sparse angle THz-CT image of the corresponding section as input to the network to preliminarily obtain a denoised reconstructed THz-CT image; according to the defined loss function, the loss of the difference value between the label image and the THz-CT reconstructed image of the corresponding section is calculated, and then the gradient value of the loss function is back-propagated multiple times to update the weights and bias parameters of the neural network; the convolutional neural network is repeatedly trained according to a preset training round until the overall training round reaches a preset value, and the weights and bias parameters of the neural network are stopped to obtain a trained neural network;
[0016] The loss function L of the neural network is composed of two parts of loss, including the perceptual loss L MSE and the regularization loss L 1 :
[0017]
[0018] in, and f i They represent the pixel-level data of the de-artifacted image generated by the neural network and the image collected at dense angles, respectively. λ is used to adjust the regularization weight of the data fidelity and sparse representation, which is 0.01.
[0019] Specifically, the number of training rounds for stopping updating network weights and bias parameters is preset to be greater than 100 times.
[0020] Preferably, the neural network adopts the Adam optimization algorithm and uses image enhancement technology, including rotation, translation and cropping, to increase training samples.
[0021] Compared with the prior art, the present invention has the following beneficial effects:
[0022] The present invention uses the UNet network as the basic model architecture to construct an algorithm network model based on sparse angle THz-CT reconstruction and artifact removal; on the basis of the UNet network model, an asymmetric convolution block composed of cascaded 3×3, 3×1, and 1×3 convolution kernels is introduced into the basic convolution block, and a channel attention mechanism of feature interconnection is used between the 3×3 kernel convolution block and the asymmetric convolution block to be closely connected together, which is conducive to spatial feature extraction from the horizontal and vertical directions and enhancing the fluidity of spatial feature information; a joint loss of perception terms and regularization terms is adopted to make the model smoother and simpler, and to make it more generalized; the reconstruction method of the present invention improves the reconstruction quality of the image while removing the artifacts of the THz-CT image, and can enhance the clarity of the edge contour and features. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 This is a flowchart of the deep learning-based sparse angle THz-CT image reconstruction method involved in the present invention.
[0024] Figure 2 The figure is a schematic diagram of a neural network model involved in the present invention.
[0025] Figure 3 The application experiment samples (a) and the three-dimensional reconstruction image experiment results (b) in Example 1 of the present invention.
[0026] Figure 4 It is a sinusoidal image of a certain section of the experimental sample in Example 1 of the present invention and a corresponding slice image reconstructed by a filtered back projection algorithm.
[0027] Figure 5 The THz-CT image (bottom) and the three-dimensional reconstructed image (top) of the same experimental sample obtained by different methods in Example 2 of the present invention. DETAILED DESCRIPTION
[0028] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0029] Example 1
[0030] See also Figure 1 The present embodiment relates to a sparse angle THz-CT image reconstruction method based on deep learning, which specifically includes the following steps:
[0031] (S1) collecting sparse angle THz-CT images and dense angle THz-CT images of training samples, wherein the sparse angle THz-CT images of the training samples in the thickness direction constitute an image set, and the dense angle THz-CT images of the training samples in the thickness direction constitute a label set, and the training set includes a label set and an image set;
[0032] (S2) inputting the training set into the neural network model for training to obtain a trained neural network model;
[0033] (S3) inputting the sparse angle THz-CT image to be reconstructed into the trained neural network model to obtain a high-resolution THz-CT reconstructed image;
[0034] (S4) The THZ-CT reconstructed image corresponding to the sparse angle THz-CT image of the entire height of the sample to be reconstructed is deeply stacked along the plane axis of the THZ-CT reconstructed image to obtain a three-dimensional image of the sample to be reconstructed.
[0035] Specifically, the sparse angle THz-CT image acquisition method of step (S1) and step (S3) is as follows: when the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample section is collected by sparse angle sampling to obtain a sinusoidal image, the sparse terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the sinusoidal image is processed by a filtered back projection algorithm to obtain a slice image of the corresponding section, that is, a sparse angle THz-CT image.
[0036] Similarly, the dense-angle THz-CT image acquisition method in step (S1) is as follows: every time the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample cross section is collected by dense-angle sampling to obtain a sinusoidal image, the dense terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the preprocessed sinusoidal image is processed by a fast shrinkage threshold soft iterative algorithm to obtain a dense-angle THz-CT image.
[0037] Specifically, the preprocessing includes filtering, truncation, multiple averaging, background denoising and spectrum power integration.
[0038] In the prior art, when extremely sparse angle THz-CT images are directly reconstructed in three dimensions, the reconstructed results obtained have serious artifact information, resulting in blurred and distorted images and low spatial resolution. The neural network reconstruction method based on deep learning in the present invention shows superior performance in the three-dimensional reconstruction of sparse angle THz-CT images, and can restore high-quality reconstructed images under extremely sparse angle sampling. Since the terahertz imaging system can usually only distinguish the smallest details at the sub-millimeter level, the results often appear smooth when imaging fine structures. The reconstruction of sparse angles must take into account the feature extraction of smooth images, and the extraction of spatial features in different directions can improve network performance.
[0039] To this end, the neural network in the present invention adopts the UNet codec structure as the basic model architecture, and introduces asymmetric convolution in the convolution block of the encoder as the basic convolution block of the reconstruction model, so that the model can enhance the ability of multi-directional feature extraction, thereby capturing richer image details and texture information; the structure of the basic convolution block of the present invention is as follows Figure 2 As shown in the figure, the convolution block is a key component of the codec, and its purpose is to achieve compression and extraction of image information. The basic convolution block is composed of a 3*3 convolution, an asymmetric convolution block composed of a cascade of 3×3, 3×1, and 1×3 convolution kernels, a linear rectification activation function, and a normalization layer.
[0040] See also Figure 2 , the channel attention mechanism is introduced into the neural network used. The attention mechanism simulates human visual processing and can adaptively allocate input weights in the network. This attention mechanism can prompt the network to pay more attention to the areas of interest in the image and improve the network's performance on image feature dependencies. Through adaptive weighting, the network can more accurately capture the location where image information is concentrated, thereby improving the network's information content and performance.
[0041] To this end, the basic convolution block in the present invention adopts 3*3 convolution, asymmetric convolution and channel attention mechanism, and introduces the channel attention mechanism in 3*3 convolution and asymmetric convolution. The channel attention mechanism consists of Extract and Transform, which interconnect adjacent features and promote the flow of spatial information, so that the network can distinguish the feature responses of different channels to different tasks, enhance the key information of multiple channels, and restore the image edges more clearly.
[0042] Example 1
[0043] This embodiment uses a specific case to illustrate the sparse angle THz-CT image reconstruction method based on a neural network of the present invention:
[0044] 1. Collection of experimental data
[0045] 1.1 Experimental setup
[0046] The THz instruments used in the experiment of the present invention are the T-Ray 5000 THz time-domain spectroscopy system of Picometrix and a self-built three-dimensional motion control console, wherein the effective spectral range of the THz equipment is 0.1 to 2.2 THz, and the peak dynamic range is greater than 70 dB.
[0047] 1.2 Data Collection
[0048] The sample is placed on a motion control platform located at the focal plane of the THz-TDS system, and a transmission imaging scan is performed. When the sample moves a set height in the thickness direction, a rotating continuous line scanning method is used to perform equal-interval scanning and sampling on the periphery of the height. In this embodiment, dense angle sampling is when the sample rotates 180° (the data of the other 180° corresponds to the collected data in reverse), and a terahertz pulse signal is collected every 5°, and a total of 36 samples are sampled. Sparse angle sampling is when the sample rotates 180°, and a terahertz pulse signal is collected every 20°, and a half-circle is rotated, and a total of 9 samples are sampled. Of course, the angles of dense sampling and sparse sampling can also be other. In the present invention, dense sampling, sufficient sampling data, and a clear cross-sectional image can be obtained through the existing fast shrinkage threshold soft iteration algorithm, and three-dimensional reconstruction can be performed directly. However, sparse sampling has less sampling data, and a clear cross-sectional image cannot be obtained, and it is difficult to directly perform three-dimensional reconstruction. The data corresponding to a section collected by rotating 180° is a group, and a group of data obtains a sinusoidal image corresponding to the section.
[0049] The sparse terahertz pulse signal of the terahertz transmission imaging system is obtained by sparse angle sampling, and the dense terahertz pulse signal of the terahertz transmission imaging system is obtained by dense angle sampling.
[0050] 1.3 Data Preprocessing
[0051] The terahertz pulse signal is preprocessed, including filtering, truncation, multiple averaging, background denoising and spectral power integration, to obtain the THz projection data of the sample, that is, to obtain the preprocessed sinusoidal image; then, the slice image is reconstructed by the existing filtered back projection algorithm or the fast shrinkage threshold soft iteration algorithm, and the slice image is a cross-sectional image of the sample after the sinusoidal image reconstruction, that is, the THz-CT image.
[0052] 1.4 Dataset Creation
[0053] This embodiment uses products made of PP, acrylic and resin materials as samples for network training, and mixes samples of different shapes of the three materials to obtain 6 training samples. The terahertz pulse signals of the entire thickness of the 6 training samples are collected according to the above-mentioned dense angle sampling method. After scanning and preprocessing, the 6 training samples obtain a total of 423 groups of THz-CT images as a data set, which includes a label set and an image set. The data set is divided into a training set, a validation set and a test set in a ratio of 4:1:1, and image enhancement technology is applied to the training set and the validation set to increase the data set samples, wherein the label set consists of slice images reconstructed by a fast shrinkage threshold soft iterative algorithm under 36 acquisition angles, and the image set consists of slice images reconstructed by a filtered back projection algorithm under 9 acquisition angles.
[0054] The training set data is used for network training, the validation set is used to correct the model parameters, the training set is used for repeated training to obtain a trained model, and then the validation set is used to preliminarily verify the quality of the model. If the verification is qualified, the model is exported and training is stopped, and the test set is used as the final evaluation standard.
[0055] The image enhancement technology includes rotation, translation and cropping to increase the training samples; in this embodiment, the image enhancement technology selects rotation and translation.
[0056] 2. Establishment and training of neural network
[0057] The neural network is an improved UNet network with a neural network structure depth of 4 layers, including enhanced convolution blocks (Enhanced Conv), basic convolution blocks (Basic Conv), pooling layers (Pooling), and upsampling layers (Upsampling). The encoder part consists of an enhanced convolution block and a pooling layer to form a downsampling module. After multiple operations, abstract features are obtained. Then the decoder part consists of an upsampling layer and a basic convolution block to form an upsampling module. After multiple operations, the target image is restored from the abstract features. Finally, a layer of 1*1 convolution is performed to reduce the dimension to 1 channel to obtain the target output. The convolution layers of corresponding dimensions in the encoder and decoder are channel-spliced. The enhanced convolution block (Enhanced Conv) of the encoder consists of a 3*3 convolution block, an asymmetric convolution (ACM Conv), a channel attention mechanism (SE Block), a linear rectification activation function, and a batch normalization layer. The basic convolution block (Basic Conv) consists of two layers of 3*3 convolution blocks, a linear rectification activation function and a batch normalization layer; the introduced channel attention mechanism is divided into two parts: Extract global feature extraction and Transform channel weighted manifold transformation, which are introduced in 3*3 convolution blocks and asymmetric convolution blocks, and the attention features of adjacent channels are interconnected; the number of downsampling channels of the overall network is 64, 128, 256, and 512, respectively, and the upsampling is the opposite.
[0058] The training set is sent to the neural network for training. The two image sets sent in are first reconstructed and denoised through the network to generate a clearer reconstructed image. At this time, according to the defined loss function, the loss of the difference value between the corresponding label image and the reconstructed image is calculated, and then the gradient value of the loss function is back-propagated multiple times to update the weights and bias parameters of the neural network; the convolutional neural network is repeatedly trained according to the preset 100 rounds, and the model with the best performance in the repeated training process is stored until the overall training round reaches the preset value, and the weights and bias parameters of the neural network are stopped to obtain a fully trained generative adversarial network.
[0059] The loss function of the neural network is composed of two parts of loss, including the perceptual loss mean square error (L MSE ) and L 1 Regularization loss:
[0060]
[0061] in, and f i They represent the pixel-level data of the de-artifacted image (reconstructed image) and the image (label image) collected from dense angles generated by the neural network, respectively. λ is used to adjust the regularization weight of the data fidelity and sparse representation, which is 0.01. L MSE The loss is used to calculate the difference between the label value and the predicted value to further optimize the output target, L 1 The regularization term is used to introduce prior information into the optimization objective to improve the robustness and generalization ability of the estimation.
[0062] See also Figure 1 The neural network adopts the Adam optimization algorithm, the training rounds are set to 100, and the network optimization learning rate is 0.001.
[0063] 3. 3D Reconstruction
[0064] This embodiment uses polymer barrels and agarwood as experimental samples for application. Figure 3 (a) shows that the diameter of the polymer barrel is 4 cm. Agarwood is placed inside the barrel. Both materials are translucent in the terahertz wave range. Agarwood is placed inside the barrel as a test sample. According to the data acquisition method, every time the sample moves a set height in the thickness direction, sparse angle sampling is performed on the test sample to obtain multiple groups of sinusoidal images collected at 9 angles (such as Figure 4 ), the sinusoidal image is reconstructed by a filtered back projection algorithm to obtain a slice image (such as Figure 4 ).
[0065] Multiple sets of slice images of the test samples obtained by sparse angle sampling are input into the trained neural network (improved UNet network structure) to obtain multiple sets of reconstructed slice images. Finally, the multiple sets of reconstructed slice images are deeply stacked along the plane axis of the image to obtain a three-dimensional image of the sample.
[0066] The specific processing process is as follows:
[0067] (1) The slice image is input into the neural network. The basic convolution layer of the encoder first uses "3×3 convolution-batch normalization-linear rectification activation function ReLu" to obtain preliminary spatial information, and then passes through the global average pooling layer as Extract to extract global spatial information. Then, it passes through a fully connected layer and activation function ReLu and a Transform block composed of a fully connected layer and activation function Sigmoid to calculate the spatial and channel attention weights through channel number compression and restoration to obtain the channel attention information G(x);
[0068] (2) The weighted features of spatial information and channel attention information are further aggregated with asymmetric convolution:
[0069] First, the asymmetric convolution is composed of a cascade of "3×1, 3×3 and 1×3" convolution kernels, followed by batch normalization and activation function ReLu; secondly, the spatial information after the asymmetric convolution is further extracted through Extract for global spatial information, and this part is interconnected with the previous channel attention weight G(x); then, the channel attention weight of this part is obtained through the Transform block and weighted with the spatial information of the asymmetric convolution.
[0070] (3) The encoder part combines the basic convolutional blocks to gradually reduce the input image x through multiple downsampling operations and extracts multi-scale feature representations to capture semantic information at different levels.
[0071] (4) Based on the encoder downsampling, the improved U-Net decoder part restores the feature map to its original size through upsampling operations, and gradually restores the detail information to obtain the final feature output. Each upsampling stage usually includes a transposed convolution layer (deconvolution layer) and a skip connection. The skip connection connects the feature map of the corresponding layer in the encoder with the feature map in the decoder to help effectively transfer low-level and high-level information.
[0072] The imaging results of the agarwood in the barrel are as follows Figure 3As shown in (b), the results show the model's ability to reconstruct complex-shaped samples at 9 angles (limited by the thickness range of the system's terahertz wave penetration, the imaging scanning range does not include the top surface and the base of the sample). The reconstructed image clearly shows the structural information of the agarwood inside the optically opaque barrel. The method of the present invention takes 1 / 4 of the acquisition time of the original label to achieve good imaging results.
[0073] Example 2
[0074] In order to verify the performance of the proposed model, we used filtered back projection (FBP) as the original input reference, simultaneous algebraic iteration (ART), fast shrinkage threshold soft iteration (FISTA) and three reconstruction methods to reconstruct THz-CT images and 3D tomographic images under 9-angle projection data, and compared them with the reconstruction results of the reconstruction model of the present invention (Proposed). The test was carried out using the test set separated from the training samples, and the number of iterations of each algorithm was set to 100 epochs. The results are shown in Figure 2. Figure 5 shown.
[0075] It can be seen from the test results that compared with the dense angle THz-CT image obtained under 36-angle projection data and the three-dimensional tomography image (Ground truth) directly constructed therefrom, the THz-CT image input with 9-angle projection data has a large number of stripe artifacts and has lost basic structural details. The image edges of the ART reconstruction result are blurred and there is a large gap with the actual thickness. The FISTA reconstruction result restores more detailed features, but because the sampling data is too sparse, the reconstructed edges are obviously incoherent and the artifacts cannot be completely removed. The model proposed in the present invention effectively suppresses artifacts and has a high degree of structural restoration, and has better reconstruction effects both in the edge contour and the texture part of the central area.
[0076] The above is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A sparse angle THz-CT image reconstruction method based on deep learning, characterized in that: The specific steps include: (S1) collecting sparse angle THz-CT images and dense angle THz-CT images of training samples, wherein the sparse angle THz-CT images of the training samples in the thickness direction constitute an image set, and the dense angle THz-CT images of the training samples in the thickness direction constitute a label set, and the training set includes a label set and an image set; (S2) inputting the training set into the neural network model for training to obtain a trained neural network model; The neural network is an improved UNet network with a 4-layer structure depth, including enhanced convolution blocks, basic convolution blocks, pooling layers, and upsampling layers. The encoder part consists of an enhanced convolution block and a pooling layer to form a downsampling module. After multiple operations, the abstract features are obtained. Then, the decoder part consists of an upsampling layer and a basic convolution block to form an upsampling module. After multiple operations, the target image is restored from the abstract features. Finally, a 1*1 convolution layer is connected to reduce the dimension to 1 channel to obtain the target output. The corresponding dimensional convolution layers in the encoder and decoder are channel-joined. The enhanced convolution block of the encoder consists of It consists of a 3*3 convolution block, an asymmetric convolution block, a channel attention mechanism, a linear rectification activation function and a batch normalization layer. The basic convolution block of the decoder consists of two layers of 3*3 convolution blocks, a linear rectification activation function and a batch normalization layer. The introduced channel attention mechanism is divided into two parts: Extract global feature extraction and Transform channel weighted manifold transformation. It is introduced in the 3*3 convolution block and the asymmetric convolution block, and the adjacent channel attention features are interconnected. The number of downsampling channels of the overall network is 64, 128, 256, and 512, respectively, and the upsampling is the opposite. (S3) inputting the sparse angle THz-CT image to be reconstructed into the trained neural network model to obtain a high-resolution THz-CT reconstructed image; (S4) performing depth stacking of the THz-CT reconstructed image corresponding to the sparse angle THz-CT image of the entire height of the sample to be reconstructed along the plane axis of the THz-CT reconstructed image to obtain a three-dimensional image of the sample to be reconstructed; In step (S2), the neural network uses the dense angle THz-CT image as the label image and the sparse angle THz-CT image of the corresponding section as input to the network to preliminarily obtain a denoised reconstructed THz-CT image; according to the defined loss function, the loss of the difference value between the label image and the THz-CT reconstructed image of the corresponding section is calculated, and then the gradient value of the loss function is back-propagated multiple times to update the weight and bias parameters of the neural network; the convolutional neural network is repeatedly trained according to a preset training round until the overall training round reaches a preset value, and the weight and bias parameters of the neural network are stopped to obtain a trained neural network; The loss function L of the neural network is composed of two parts of loss, including the perceptual loss L MSE And the regularization loss L1: in, and f i They represent the pixel-level data of the de-artifacted image generated by the neural network and the image collected at dense angles, respectively. λ is used to adjust the regularization weight of the data fidelity and sparse representation, which is 0.
01.
2. The deep learning-based sparse angle THz-CT image reconstruction method according to claim 1, characterized in that: The sparse angle THz-CT image acquisition method of step (S1) and step (S3) is as follows: when the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample section is collected by sparse angle sampling to obtain a sinusoidal image, the sparse terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the sinusoidal image is processed by a filtered back projection algorithm to obtain a slice image of the corresponding section, that is, a sparse angle THz-CT image. Similarly, the dense-angle THz-CT image acquisition method in step (S1) is as follows: every time the sample moves a set height in the thickness direction, the terahertz pulse signal of one circle of the sample cross section is collected through dense-angle sampling, the dense terahertz pulse signal of the terahertz transmission imaging system is preprocessed to obtain a preprocessed sinusoidal image, and the preprocessed sinusoidal image is processed by a fast shrinkage threshold soft iterative algorithm to obtain a dense-angle THz-CT image.
3. The deep learning-based sparse angle THz-CT image reconstruction method according to claim 2, characterized in that: The preprocessing includes filtering, truncation, multiple averaging, background denoising and spectrum power integration.
4. The deep learning-based sparse angle THz-CT image reconstruction method according to claim 1, characterized in that: The number of training rounds for stopping updating network weights and bias parameters is preset to be greater than 100.
5. The deep learning-based sparse angle THz-CT image reconstruction method according to claim 1, characterized in that: The neural network adopts the Adam optimization algorithm and uses image enhancement technology, including rotation, translation and cropping, to increase training samples.
