Target reconstruction method for sparse view angle image

Abstract

The invention relates to a target reconstruction method for a sparse view angle image, belongs to the technical field of image reconstruction, and solves the problems of low quality and artifacts of a reconstructed image obtained by reconstruction based on the sparse view angle image in the prior art. The target reconstruction method comprises the following steps: detecting and collecting a target object through an X-ray machine to obtain a detection projection image with a small number of projection visual angles; based on a linear interpolation method, performing interpolation on the detection projection image in the projection visual angle direction to obtain an interpolation recovery projection image; inputting the interpolation recovery projection image into a pre-trained projection domain network model to obtain an optimized projection image; and reconstructing the optimized projection image based on a preset image reconstruction algorithm to obtain a target reconstructed image. And the quality of the image reconstructed by the sparse view angle image is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image reconstruction, and in particular to a target reconstruction method for sparse perspective images. Background Art

[0002] At present, X-ray machines, as a non-invasive inspection equipment, have been widely used in various security inspection scenarios, such as airports, ports, hospitals and scenic spots.

[0003] In order to apply X-ray machines to complex application scenarios, usually only target images from a small number of viewing angles can be collected. When only target images from a small number of viewing angles can be obtained, the target CT images reconstructed using existing technologies have defects such as artifacts and low image quality.

[0004] Therefore, a new image reconstruction method is urgently needed. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a target reconstruction method for sparse perspective images, so as to solve the problems of low quality and artifacts of reconstructed images obtained by reconstruction based on sparse perspective images in the prior art.

[0006] An embodiment of the present invention provides a method for object reconstruction for a sparse perspective image, the method comprising:

[0007] The target object is detected and collected by an X-ray machine to obtain a detection projection image with a small number of projection viewing angles;

[0008] Based on the linear interpolation method, the detection projection image is interpolated in the projection viewing angle direction to obtain the interpolated restored projection image;

[0009] The interpolated restored projection image is input into the pre-trained projection domain network model to obtain the optimized projection image;

[0010] The optimized projection image is reconstructed based on a preset image reconstruction algorithm to obtain a target reconstructed image.

[0011] Based on a further improvement of the above method, the interpolation of the detected projection image in the projection viewing angle direction includes:

[0012] Determine one or more interpolation perspectives to be inserted between two adjacent projection perspectives, and determine the pixel points under each interpolation perspective;

[0013] Determine an interpolation pixel value of a pixel point under each interpolation viewing angle according to the interpolation viewing angle of the pixel point under each interpolation viewing angle, two adjacent projection viewing angles, and two adjacent projection pixel values ​​in the projection viewing angle direction;

[0014] The pixel points of all projection view angles and all interpolation view angles are arranged in the order of projection view angles to obtain the interpolated restored projection image.

[0015] Based on the further improvement of the above method, the interpolated pixel value of each pixel under the interpolation perspective is calculated by the following formula:

[0016]

[0017] Among them, β k and β k+1 Represents two adjacent projection perspectives, β k ′ represents the interpolation perspective, P(β k′ ,i) represents the interpolation viewing angle β k′ The interpolated pixel value of the i-th pixel under k , i) and P(β k+1 , i) represent the projection viewing angle β k and β k+1 The projected pixel value of the i-th pixel under .

[0018] Based on the further improvement of the above method, the projection domain network model includes a feature extraction module, an upsampling module and a splicing module connected in sequence;

[0019] The feature extraction module is used to extract multi-scale features from the input interpolated and restored projection image, and input the obtained feature maps of multiple scales into the upsampling module:

[0020] An upsampling module is used to perform layer-by-layer upsampling and splicing of multiple feature maps of different scales to obtain a tenth feature map, and output the obtained tenth feature map to the splicing module;

[0021] The splicing module is used to splice the input interpolated restored projection image and the tenth feature map, and use the spliced ​​feature map as the optimized projection image.

[0022] Based on a further improvement of the above method, the feature extraction module includes a first deformable convolution module, a preset number of first dual multi-scale deformable convolution modules and a second dual multi-scale deformable convolution module connected in sequence;

[0023] A first deformable convolution module is used to perform a deformable convolution operation and a multi-scale deformable convolution operation on the input interpolated restored projection image to obtain a first feature map;

[0024] A first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the first feature map to obtain a second feature map;

[0025] Multiple intermediate first dual multi-scale deformable convolution modules respectively perform dual multi-scale deformable convolution operations on the feature maps output by the previous first dual multi-scale deformable convolution module to obtain feature maps of different scales as multiple third feature maps;

[0026] The last first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the feature map output by the previous first dual multi-scale deformable convolution module to obtain a fourth feature map;

[0027] The second dual multi-scale deformable convolution module is used to perform a dual multi-scale deformable convolution operation on the fourth feature map to obtain a fifth feature map.

[0028] Based on a further improvement of the above method, the first deformable convolution module includes a deformable convolution module and a multi-scale deformable convolution module connected in sequence;

[0029] The first dual multi-scale deformable convolution module and the second dual multi-scale deformable convolution module have the same structure, and both include two multi-scale deformable convolution modules with the same structure;

[0030] The multi-scale deformable convolution module performs multi-scale deformable convolution operations, batch normalization operations, and ReLU activation function operations on the input feature map.

[0031] Based on the further improvement of the above method, the multi-scale deformable convolution operation includes the following steps:

[0032] Perform channel separation on the input feature map to obtain feature maps of multiple different channels;

[0033] The feature maps of multiple different channels are subjected to convolution operations of different scales to obtain feature maps of multiple different scales;

[0034] Channel splicing is performed on multiple feature maps of different scales, and the obtained feature map is spliced ​​with the input feature map as the output feature map.

[0035] Based on the further improvement of the above method, the upsampling module includes a two-dimensional deconvolution module, a preset number of deformable convolution and deconvolution modules and a second deformable convolution module;

[0036] The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the input fifth feature map to obtain a sixth feature map;

[0037] The sixth feature map is concatenated with the fourth feature map of the same scale and input into the first deformable convolution and deconvolution module. Each deformable convolution and deconvolution module is used to perform a deformable convolution operation and a two-dimensional deconvolution operation on the input feature map, and output feature maps of different scales.

[0038] The input of each deformable convolution and deconvolution module is the feature map obtained by concatenating the feature map output by the previous deformable convolution and deconvolution module and the third feature map of the same scale output by the feature extraction module;

[0039] The input of the last deformable convolution and deconvolution module is the concatenation of the feature map output by the previous deformable convolution and deconvolution module and the second feature map of the same scale;

[0040] The feature map output by the last deformable convolution and deconvolution module is concatenated with the first feature map of the same scale and input into the second deformable convolution module to obtain the tenth feature map.

[0041] Based on the further improvement of the above method, the deformable convolution and deconvolution module includes a deformable convolution module and a two-dimensional deconvolution module connected in sequence;

[0042] The deformable convolution module performs deformable convolution, batch normalization, and ReLU activation function operations on the input feature map in sequence;

[0043] The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the input feature map.

[0044] Based on a further improvement of the above method, the second deformable convolution module includes a deformable convolution module and a common convolution module connected in sequence;

[0045] The ordinary convolution module performs convolution operation, batch normalization operation and ReLU activation function operation on the input feature map to obtain the tenth feature map.

[0046] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0047] 1. By performing linear interpolation on the detected projection image to obtain the interpolated restored projection image, the sampling density of the projection data is supplemented, reducing the impact of insufficient sampling density caused by sparse view images. The interpolated restored projection image is optimized by combining the projection domain network model, significantly reducing the error caused by linear interpolation and improving the quality of the reconstructed image.

[0048] 2. When performing linear interpolation on the detection projection image to increase the sampling density, the pixel value under the interpolation perspective is determined by using the interpolation perspective, the adjacent projection perspective, and the pixel value corresponding to the projection perspective, so that the interpolation is more natural and the error is smaller;

[0049] 3. In the projection domain network model, the interpolated restored projection image is feature extracted and predicted through the sequentially connected feature extraction module, upsampling module and splicing module to obtain the optimized projection image, which reduces the error of the interpolated restored projection image and improves the accuracy of the reconstructed image.

[0050] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0052] Figure 1 A schematic flow chart of a method for object reconstruction of sparse perspective images provided by an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of a detection projection image with a small number of projection viewing angles provided by an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of interpolating and restoring a projected image according to an embodiment of the present invention;

[0055] Figure 4 A schematic diagram of an optimized projection image provided by an embodiment of the present invention;

[0056] Figure 5 A schematic diagram of a target reconstructed image provided by an embodiment of the present invention;

[0057] Figure 6 A schematic diagram of the structure of a projection domain network model provided by an embodiment of the present invention;

[0058] Figure 7 A schematic structural diagram of a first deformable convolution module provided in an embodiment of the present invention;

[0059] Figure 8 A schematic diagram of the process of a multi-scale deformable convolution operation provided by an embodiment of the present invention;

[0060] Figure 9 A schematic diagram of the structure of a deformable convolution and deconvolution module provided in an embodiment of the present invention;

[0061] Figure 10 A schematic structural diagram of a second deformable convolution module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0063] A specific embodiment of the present invention discloses a method for object reconstruction of sparse perspective images, such as Figure 1 As shown, the target reconstruction method includes:

[0064] Step S1: detecting and collecting the target object by an X-ray machine to obtain a detection projection image with a small number of projection viewing angles;

[0065] Step S2: Based on the linear interpolation method, the detected projection image is interpolated in the projection viewing angle direction to obtain an interpolated restored projection image;

[0066] Step S3: inputting the interpolated restored projection image into the pre-trained projection domain network model to obtain the optimized projection image;

[0067] Step S4: reconstructing the optimized projection image based on a preset image reconstruction algorithm to obtain a target reconstructed image.

[0068] Specifically, an X-ray machine is a device that uses X-rays to penetrate tissues of different densities to produce absorption differences to form images. The target object is placed at the detection point of the X-ray machine, and images of the target object under several projection angles are collected. For example, the target object is sampled every 2° to obtain a projection angle image. The intervals between the angles are large, which will lead to undersampling problems, and then cause aliasing artifacts in the reconstructed image.

[0069] It is understandable that when it is found that the amount of detection projection image data obtained by sampling with an X-ray machine is small, a higher quality reconstructed image can be obtained by the reconstruction method provided by the present invention.

[0070] Specifically, such as Figure 1 As shown, in step S1, a detection projection image with a small number of projection viewing angles is obtained. Specifically, the X-ray machine can capture an image at each projection viewing angle, and all the captured images are arranged and spliced ​​in the order of projection viewing angles to obtain the following: Figure 2 The detection projection image shown, where Figure 2 The horizontal direction is the detector direction, and the vertical direction is the projection viewing angle direction. Figure 2 Each row in can be understood as the sampled image data at each projection perspective.

[0071] Specifically, such as Figure 1 As shown, in step S2, the detection projection image is linearly interpolated in the projection viewing angle direction to obtain an interpolated restored projection image, as shown in FIG. Figure 3As shown, the resolution of the interpolated restored projection image in the projection viewing angle direction is improved, and the sampling data is improved.

[0072] Preferably, the interpolating the detected projection image in the projection viewing angle direction includes:

[0073] Determine one or more interpolation perspectives to be inserted between two adjacent projection perspectives, and determine the pixel points under each interpolation perspective;

[0074] Determine an interpolation pixel value of a pixel point under each interpolation viewing angle according to the interpolation viewing angle of the pixel point under each interpolation viewing angle, two adjacent projection viewing angles, and two adjacent projection pixel values ​​in the projection viewing angle direction;

[0075] The pixel points of all projection view angles and all interpolation view angles are arranged in the order of projection view angles to obtain the interpolated restored projection image.

[0076] Specifically, when performing interpolation, one or more interpolation perspectives can be set according to the size of the gap between two adjacent projection perspectives. It can be understood that the gap between any two adjacent projection perspectives may be different, so the number of interpolation perspectives required to be inserted between any two adjacent projection perspectives will be different.

[0077] Specifically, the number of pixels at each projection viewing angle is the same, and when linear interpolation is performed, the number of pixels at each interpolation viewing angle is made the same as the number of pixels at the projection viewing angle.

[0078] When calculating the interpolation pixel value of each pixel point under each interpolation viewing angle, two adjacent projection viewing angles and two corresponding projection pixel values ​​in the projection viewing angle direction are determined. Figure 2 As shown, the first row represents the projection pixel values ​​under a 2° projection angle of view, each row includes 1000 pixels, and the second row represents the projection pixel values ​​under a 4° projection angle of view, each row includes 1000 pixels. Now, an interpolation angle needs to be inserted between the first and second rows. In this case, the interpolation angle of view can be set to 3°, so a row of pixels, i.e., 1000 rows, needs to be inserted under the 3° interpolation angle of view. When calculating the interpolation pixel values ​​of any column of pixels under the 3° interpolation angle of view, the interpolation pixel values ​​can be determined based on the projection pixel values ​​of the pixels of any column under the 2° projection angle of view and the projection pixel values ​​of the pixels of any column under the 4° projection angle of view.

[0079] Preferably, the interpolated pixel value of each pixel under the interpolation viewing angle is calculated by the following formula:

[0080]

[0081] Among them, β k and β k+1 Represents two adjacent projection perspectives, βk′ represents the interpolation perspective, P(β k′ , i) represents the interpolation viewing angle β k′ The interpolated pixel value of the i-th pixel under k ,i) and P(β k+1 ,i) respectively represent the projection viewing angle β k and β k+1 The projected pixel value of the i-th pixel under .

[0082] Specifically, the above formula calculates the interpolated pixel value of any pixel point under the interpolated perspective by the angle difference between the interpolated perspective and two adjacent projection perspectives, so that the interpolated pixel point is closer to the real pixel value.

[0083] Specifically, such as Figure 3 As shown, after the interpolation pixel values ​​of all pixels under all interpolation viewing angles are calculated, the pixel points of all projection viewing angles and all interpolation viewing angles are arranged in the order of projection viewing angles to obtain the interpolated restored projection image.

[0084] Specifically, such as Figure 1 As shown, in step S3, the interpolated restored projection image obtained in step S2 is input into the pre-trained projection domain network model, and the optimized projection image is predicted by the projection domain network model, as shown in FIG. Figure 4 As shown, the optimized projection image and the interpolated restored projection image have the same size, the difference is that the optimized projection image is an optimized interpolated restored projection image, which can correct and optimize the interpolated pixels in the interpolated restored projection image. The optimized projection image after correction is closer to the target object. In step S4, the optimized projection image is reconstructed by a preset image reconstruction algorithm, and the image quality of the obtained target reconstructed image is better.

[0085] Preferably, the preset image reconstruction algorithm can use a filtered back projection algorithm, an algebraic reconstruction algorithm or a maximum likelihood expectation maximum algorithm to reconstruct the optimized projection image to obtain a target reconstructed image, such as Figure 5 shown.

[0086] Preferably, if Figure 6 As shown, the projection domain network model includes a feature extraction module, an upsampling module and a splicing module connected in sequence;

[0087] The feature extraction module is used to extract multi-scale features from the input interpolated and restored projection image, and input the obtained feature maps of multiple scales into the upsampling module:

[0088] An upsampling module is used to perform layer-by-layer upsampling and splicing of multiple feature maps of different scales to obtain a tenth feature map, and output the obtained tenth feature map to the splicing module;

[0089] The splicing module is used to splice the input interpolated restored projection image and the tenth feature map, and use the spliced ​​feature map as the optimized projection image.

[0090] Specifically, such as Figure 6 As shown, the projection domain network model includes a feature extraction module, an upsampling module, and a splicing module. The feature extraction module is used to receive the interpolated restored projection image and perform multi-scale feature extraction on the interpolated restored projection image to obtain multiple first to fifth feature maps of different scales, which are respectively input into the upsampling module; the upsampling module performs layer-by-layer upsampling and splicing on the multiple first to fifth feature maps of different scales, and outputs the obtained tenth feature map P10 to the splicing module; the splicing module is used to splice the interpolated restored projection image and the tenth feature map P10, and use the spliced ​​feature map as the optimized projection image.

[0091] Preferably, if Figure 6 As shown, the feature extraction module includes a first deformable convolution module, a preset number of first dual multi-scale deformable convolution modules and a second dual multi-scale deformable convolution module connected in sequence;

[0092] A first deformable convolution module is used to perform a deformable convolution operation and a multi-scale deformable convolution operation on the input interpolated restored projection image to obtain a first feature map;

[0093] A first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the first feature map to obtain a second feature map;

[0094] Multiple intermediate first dual multi-scale deformable convolution modules respectively perform dual multi-scale deformable convolution operations on the feature maps output by the previous first dual multi-scale deformable convolution module to obtain feature maps of different scales as multiple third feature maps;

[0095] The last first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the feature map output by the previous first dual multi-scale deformable convolution module to obtain a fourth feature map;

[0096] The second dual multi-scale deformable convolution module is used to perform a dual multi-scale deformable convolution operation on the fourth feature map to obtain a fifth feature map.

[0097] Specifically, such as Figure 6As shown, after the interpolated restored projected image is input into the feature extraction module, the obtained feature map passes through the first deformable convolution module, a preset number of first dual multi-scale deformable convolution modules and the second dual multi-scale deformable convolution module in sequence to obtain the first feature map P1, the second feature map P2, multiple third feature maps P3, the fourth feature map P4 and the fifth feature map P5 respectively.

[0098] It is worth noting that the preset number is set in advance, and the preset number is greater than or equal to 3, that is, the feature extraction module includes at least 3 first dual multi-scale deformable convolution modules.

[0099] Specifically, such as Figure 6 As shown, the preset number of first dual multi-scale deformable convolution modules includes a first first dual multi-scale deformable convolution module, multiple intermediate first dual multi-scale deformable convolution modules and a last first dual multi-scale deformable convolution module connected in sequence, wherein the feature map output by the first first dual multi-scale deformable convolution module is used as the second feature map P2, the feature map output by the intermediate first dual multi-scale deformable convolution module is used as the third feature map P3, and the feature map output by the last first dual multi-scale deformable convolution module is used as the fourth feature map P4.

[0100] It is worth noting that when the feature extraction module includes three or more first dual multi-scale deformable convolution modules, the feature extraction module outputs multiple third feature maps P3. It can be understood that the sizes of the multiple third feature maps P3 are different.

[0101] Preferably, if Figure 7 As shown, the first deformable convolution module includes a deformable convolution module and a multi-scale deformable convolution module connected in sequence;

[0102] The first dual multi-scale deformable convolution module and the second dual multi-scale deformable convolution module have the same structure, and both include two multi-scale deformable convolution modules with the same structure;

[0103] The multi-scale deformable convolution module performs multi-scale deformable convolution operations, batch normalization operations, and ReLU activation function operations on the input feature map.

[0104] Specifically, the first dual multi-scale deformable convolution module and the second dual multi-scale deformable convolution module have the same structure, and both include two multi-scale deformable convolution modules with the same structure connected in sequence.

[0105] Specifically, the deformable convolution module performs a deformable convolution operation on the input feature map, and the multi-scale deformable convolution module performs a multi-scale deformable convolution operation, a batch normalization operation, and a ReLU activation function operation on the input feature map.

[0106] Specifically, such as Figure 6 As shown, the first deformable convolution module performs a deformable convolution operation and a multi-scale deformable convolution operation on the input interpolated restored projection image to obtain a first feature map P1.

[0107] Preferably, the multi-scale deformable convolution operation includes the following steps:

[0108] Perform channel separation on the input feature map to obtain feature maps of multiple different channels;

[0109] The feature maps of multiple different channels are subjected to convolution operations of different scales to obtain feature maps of multiple different scales;

[0110] Channel splicing is performed on multiple feature maps of different scales, and the obtained feature map is spliced ​​with the input feature map as the output feature map.

[0111] Specifically, such as Figure 8 As shown in the figure, the input feature map is firstly channel-separated in the multi-scale deformable convolution operation, and the input feature map is divided into feature maps of multiple different channels. Secondly, the feature maps of multiple different channels are subjected to convolution operations of different scales, such as 3×3 deformable convolution operation, 1×11 deformable convolution operation, 11×1 deformable convolution operation and identity mapping operation to obtain feature maps of multiple scales. Finally, the feature maps of multiple scales are spliced ​​through channels and output as the output feature map.

[0112] Preferably, if Figure 6 As shown, the upsampling module includes a two-dimensional deconvolution module, a preset number of deformable convolution and deconvolution modules, and a second deformable convolution module;

[0113] The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the input fifth feature map to obtain a sixth feature map;

[0114] The sixth feature map is concatenated with the fourth feature map of the same scale and input into the first deformable convolution and deconvolution module. Each deformable convolution and deconvolution module is used to perform a deformable convolution operation and a two-dimensional deconvolution operation on the input feature map, and output feature maps of different scales.

[0115] The input of each deformable convolution and deconvolution module is the feature map obtained by concatenating the feature map output by the previous deformable convolution and deconvolution module and the third feature map of the same scale output by the feature extraction module;

[0116] The input of the last deformable convolution and deconvolution module is the concatenation of the feature map output by the previous deformable convolution and deconvolution module and the second feature map of the same scale;

[0117] The feature map output by the last deformable convolution and deconvolution module is concatenated with the first feature map of the same scale and input into the second deformable convolution module to obtain the tenth feature map.

[0118] Specifically, the upsampling module includes a two-dimensional deconvolution module, a preset number of deformable convolution and deconvolution modules, and a second deformable convolution module. The preset number is set in advance, and the number of the first dual multi-scale deformable convolution module and the deformable convolution and deconvolution module needs to be the same.

[0119] like Figure 6 As shown, the fifth feature map P5 undergoes a two-dimensional deconvolution operation in the two-dimensional deconvolution module to obtain a sixth feature map P6. The sixth feature map P6 and the fourth feature map P4 are spliced ​​and input into the first deformable convolution and deconvolution module. The first deformable convolution and deconvolution module performs a deformable convolution operation and a two-dimensional deconvolution operation on the input feature map to obtain the seventh feature map P7; the seventh feature map P7 and the third feature map P3 of the same scale are spliced ​​and input into the middle deformable convolution and deconvolution module. The feature map output by the middle deformable convolution and deconvolution module is used as the eighth feature map P8. The feature map output by the last deformable convolution and deconvolution module is used as the ninth feature map P9. The ninth feature map P9 and the first feature map P1 are spliced ​​and input into the second deformable convolution module to obtain the tenth feature map P10.

[0120] Preferably, if Figure 9 As shown, the deformable convolution and deconvolution module includes a deformable convolution module and a two-dimensional deconvolution module connected in sequence;

[0121] The deformable convolution module performs deformable convolution, batch normalization, and ReLU activation function operations on the input feature map in sequence;

[0122] The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the input feature map.

[0123] Specifically, such as Figure 9 As shown in the figure, the feature map input to the deformable convolution and deconvolution module passes through the deformable convolution module and the two-dimensional deconvolution module in sequence, and performs deformable convolution operation, batch normalization operation and ReLU activation function operation, as well as two-dimensional deconvolution operation respectively to obtain the output feature map.

[0124] Preferably, if Figure 10 As shown, the second deformable convolution module includes a deformable convolution module and a common convolution module connected in sequence;

[0125] The ordinary convolution module performs convolution operation, batch normalization operation and ReLU activation function operation on the input feature map to obtain the tenth feature map.

[0126] Specifically, such as Figure 10 As shown, the feature map input into the second deformable convolution module passes through the deformable convolution module and the ordinary convolution module in sequence, and performs deformable convolution operation, batch normalization operation and ReLU activation function operation, as well as convolution operation, batch normalization operation and ReLU activation function operation respectively to obtain the output feature map as the tenth feature map P10.

[0127] It can be understood that the interpolated restored projection image input into the feature extraction module undergoes feature extraction in sequence, and its receptive field can dynamically adapt to changes in local projection features, thereby improving the modeling ability of the projection domain network model in irregular projections and obtaining higher quality images; at the same time, the feature maps of each scale obtained in the feature extraction module are spliced ​​with the feature maps of the same scale generated in the upsampling module, and then operated again, which can not only retain the original information but also enhance the feature expression ability, so that the obtained optimized projection image can significantly reduce the pixel value error caused by linear interpolation compared to the interpolated restored projection image.

[0128] Compared with the prior art, an embodiment of the present invention provides a target reconstruction method for sparse perspective images. By performing linear interpolation on the detection projection image, an interpolated restored projection image is obtained to supplement the sampling density of the projection data, thereby reducing the impact of insufficient sampling density caused by the sparse perspective image. The interpolated restored projection image is optimized in combination with a projection domain network model, significantly reducing the error caused by linear interpolation and improving the quality of the reconstructed image. At the same time, when linear interpolation is performed on the detection projection image to increase the sampling density, the pixel value under the interpolated perspective is determined using the interpolated perspective, the adjacent projection perspective, and the pixel value corresponding to the projection perspective, making the interpolation more natural and the error smaller. Moreover, in the projection domain network model, the interpolated restored projection image is feature extracted and predicted through a sequentially connected feature extraction module, an upsampling module, and a splicing module to obtain an optimized projection image, thereby reducing the error of the interpolated restored projection image and improving the accuracy of the reconstructed image.

[0129] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0130] The above description is only a preferred specific 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 thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for object reconstruction of sparse view images, characterized in that: The target reconstruction method comprises: The target object is detected and collected by an X-ray machine to obtain a detection projection image with a small number of projection viewing angles; Based on the linear interpolation method, the detection projection image is interpolated in the projection viewing angle direction to obtain the interpolated restored projection image; The interpolated restored projection image is input into the pre-trained projection domain network model to obtain the optimized projection image; The optimized projection image is reconstructed based on a preset image reconstruction algorithm to obtain a target reconstructed image.

2. The target reconstruction method according to claim 1, characterized in that: The interpolating the detected projection image in the projection viewing angle direction includes: Determine one or more interpolation perspectives to be inserted between two adjacent projection perspectives, and determine the pixel points under each interpolation perspective; Determine an interpolation pixel value of a pixel point under each interpolation viewing angle according to the interpolation viewing angle of the pixel point under each interpolation viewing angle, two adjacent projection viewing angles, and two adjacent projection pixel values ​​in the projection viewing angle direction; The pixel points of all projection view angles and all interpolation view angles are arranged in the order of projection view angles to obtain the interpolated restored projection image.

3. The target reconstruction method according to claim 2, characterized in that: The interpolated pixel value of each pixel under the interpolation perspective is calculated using the following formula: Among them, β k and β k+1 Represents two adjacent projection perspectives, β k′ represents the interpolation perspective, P(β k′ , i) represents the interpolation viewing angle β k′ The interpolated pixel value of the i-th pixel under k ,i) and P(β k+1 ,i) respectively represent the projection viewing angle β k and β k+1 The projected pixel value of the i-th pixel under .

4. The target reconstruction method according to claim 1, characterized in that: The projection domain network model includes a feature extraction module, an upsampling module and a splicing module connected in sequence; The feature extraction module is used to extract multi-scale features from the input interpolated and restored projection image, and input the obtained feature maps of multiple scales into the upsampling module: An upsampling module is used to perform layer-by-layer upsampling and splicing of multiple feature maps of different scales to obtain a tenth feature map, and output the obtained tenth feature map to the splicing module; The splicing module is used to splice the input interpolated restored projection image and the tenth feature map, and use the spliced ​​feature map as the optimized projection image.

5. The target reconstruction method according to claim 4, characterized in that: The feature extraction module includes a first deformable convolution module, a preset number of first dual multi-scale deformable convolution modules and a second dual multi-scale deformable convolution module connected in sequence; A first deformable convolution module is used to perform a deformable convolution operation and a multi-scale deformable convolution operation on the input interpolated restored projection image to obtain a first feature map; A first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the first feature map to obtain a second feature map; Multiple intermediate first dual multi-scale deformable convolution modules respectively perform dual multi-scale deformable convolution operations on the feature maps output by the previous first dual multi-scale deformable convolution module to obtain feature maps of different scales as multiple third feature maps; The last first dual multi-scale deformable convolution module performs a dual multi-scale deformable convolution operation on the feature map output by the previous first dual multi-scale deformable convolution module to obtain a fourth feature map; The second dual multi-scale deformable convolution module is used to perform a dual multi-scale deformable convolution operation on the fourth feature map to obtain a fifth feature map.

6. The target reconstruction method according to claim 5, characterized in that: The first deformable convolution module includes a deformable convolution module and a multi-scale deformable convolution module connected in sequence; The first dual multi-scale deformable convolution module and the second dual multi-scale deformable convolution module have the same structure, and both include two multi-scale deformable convolution modules with the same structure; The multi-scale deformable convolution module performs multi-scale deformable convolution operations, batch normalization operations, and ReLU activation function operations on the input feature map.

7. The target reconstruction method according to claim 6, characterized in that: The multi-scale deformable convolution operation includes the following steps: Perform channel separation on the input feature map to obtain feature maps of multiple different channels; The feature maps of multiple different channels are subjected to convolution operations of different scales to obtain feature maps of multiple different scales; Channel splicing is performed on multiple feature maps of different scales, and the obtained feature map is spliced ​​with the input feature map as the output feature map.

8. The target reconstruction method according to claim 5, characterized in that: The upsampling module includes a two-dimensional deconvolution module, a preset number of deformable convolution and deconvolution modules, and a second deformable convolution module; The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the fifth feature map input to obtain a sixth feature map; The sixth feature map is concatenated with the fourth feature map of the same scale and input into the first deformable convolution and deconvolution module. Each deformable convolution and deconvolution module is used to perform a deformable convolution operation and a two-dimensional deconvolution operation on the input feature map, and output feature maps of different scales. The input of each deformable convolution and deconvolution module is the feature map obtained by concatenating the feature map output by the previous deformable convolution and deconvolution module and the third feature map of the same scale output by the feature extraction module; The input of the last deformable convolution and deconvolution module is the concatenation of the feature map output by the previous deformable convolution and deconvolution module and the second feature map of the same scale; The feature map output by the last deformable convolution and deconvolution module is concatenated with the first feature map of the same scale and input into the second deformable convolution module to obtain the tenth feature map.

9. The target reconstruction method according to claim 8, characterized in that: The deformable convolution and deconvolution module includes a deformable convolution module and a two-dimensional deconvolution module connected in sequence; The deformable convolution module performs deformable convolution, batch normalization, and ReLU activation function operations on the input feature map in sequence; The two-dimensional deconvolution module performs a two-dimensional deconvolution operation on the input feature map.

10. The target reconstruction method according to claim 9, characterized in that: The second deformable convolution module includes a deformable convolution module and a common convolution module connected in sequence; The ordinary convolution module performs convolution operation, batch normalization operation and ReLU activation function operation on the input feature map to obtain the tenth feature map.

Images