A target object segmentation method and system for security CT three-dimensional images

Abstract

The present application relates to a kind of target object segmentation method and system for security CT three-dimensional image, belong to security image technical field, solve the problems, such as high computational complexity, low computational efficiency and low segmentation accuracy in prior art security CT image segmentation method, it is difficult to apply in practice.The method comprises the following steps: obtaining the original three-dimensional body data of security CT image, slicing the original three-dimensional body data to obtain a plurality of two-dimensional slice images;Based on a plurality of two-dimensional slice images, obtain each two-dimensional segmentation image based on target object, then fuse each two-dimensional segmentation image to obtain three-dimensional segmentation body data;The target object in the three-dimensional segmentation body data is filled by anchor box, and anchor box mark is added to original three-dimensional body data;Based on the original three-dimensional body data after marking, three-dimensional body data image with target object mark is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of security inspection images, and particularly relates to a target object segmentation method and system for security inspection CT three-dimensional images. BACKGROUND

[0002] Automatic target segmentation in three-dimensional X-ray computed tomography images (CT images) can segment the target objects detected in security inspection, and the segmentation of the target objects can be accurate to the boundary, which plays an important role in dangerous goods inspection in aviation baggage security inspection and has been widely concerned in the industry in recent years. Three-dimensional baggage CT scanning images (security inspection CT images) adopt pseudo-color, and there are many types of baggage images. Existing target segmentation algorithms simplify the target segmentation problem into a template-based matching problem, and the target is to determine whether the template object appears in the test baggage. This kind of algorithm proposes 3D SIFT descriptor to describe the key points in the template and the test volume, and then applies the RANSAC algorithm to select the key point matching to determine whether the template is included in the test volume data; further comparison uses different key point descriptors (including density histogram, density gradient histogram, crack and 3D SIFT) to compare the performance of template matching. However, the template-based matching method requires that the candidate object in the test baggage be exactly the same as the template object, and even a slight change in shape may lead to mismatching, and application of this kind of algorithm in aviation baggage security inspection may pose a serious threat to aviation safety.

[0003] At present, in order to improve the target classification accuracy of 3D baggage CT images, extensive research has been conducted, and it is proposed to apply an accurate 3D classifier in a sliding window method, to extract the density histogram and density gradient histogram of each key point, and to mark it with a visual unit, that is, to encode; support vector machines are used for further classification of the encoded image pixels; and random forests are used as a feature encoding method to improve the classification accuracy, which is similar to a visual cortex model to extract three-dimensional features, similar to a deep neural network, but the core of each layer is manually labeled. Most of the work of this kind of method is to use manually labeled features, and its performance is limited by human prior knowledge, and the segmentation accuracy is low, and even if these classifiers can be combined into a sliding window method to segment target objects, due to the high computational cost on each 3D image, low computational efficiency and high computational complexity, it cannot be applied in practice.

[0004] Therefore, the existing security inspection CT image segmentation method has high computational complexity, low computational efficiency and low segmentation accuracy, and is difficult to apply in practice. SUMMARY

[0005] In view of the above analysis, the embodiments of the present application aim to provide a target object segmentation method and system for security CT three-dimensional images, to solve the problems of high computational complexity, low computational efficiency and low segmentation accuracy of existing security CT image segmentation methods, which are difficult to apply in practice.

[0006] In one aspect, the embodiments of the present application provide a target object segmentation method for security CT three-dimensional images, comprising the following steps:

[0007] Obtaining the original three-dimensional body data of the security CT image, slicing the original three-dimensional body data to obtain a plurality of two-dimensional slice images;

[0008] Based on a plurality of two-dimensional slice images, obtaining each two-dimensional segmentation image based on a target object, and then fusing each two-dimensional segmentation image to obtain three-dimensional segmentation body data;

[0009] Filling the target object in the three-dimensional segmentation body data with an anchor box, and adding an anchor box label to the original three-dimensional body data;

[0010] Based on the labeled original three-dimensional body data, obtaining a three-dimensional body data image with a target object label.

[0011] Further, a plurality of two-dimensional slice images are obtained by performing the following operations:

[0012] According to the three-dimensional Cartesian coordinate system, the original three-dimensional body data is cut into slice body data of a fixed thickness along the X, Y and Z directions respectively, and is sequentially numbered; the numbering includes the cutting direction and the cutting sequence number to which the slice body data belongs;

[0013] Projecting each slice body data along the thickness direction to obtain a plurality of two-dimensional slice images, wherein the pixel value of each pixel point of the two-dimensional slice image is the maximum value of the pixel values of the voxels at the corresponding position of the slice body data in the thickness direction.

[0014] Further, by inputting a plurality of two-dimensional slice images into a 2D UNet network, each two-dimensional slice image is segmented based on a target object category to obtain each two-dimensional segmentation image based on a target object; wherein the value of each pixel point in the two-dimensional segmentation image is a category value or a background value, and the category value represents the importance level of the target object.

[0015] Further, each two-dimensional segmentation image is fused by performing the following steps to obtain three-dimensional segmentation body data:

[0016] The two-dimensional segmentation images of each cutting direction are processed as follows to obtain the corresponding segmentation body data:

[0017] arranging the two-dimensional segmentation images of the cutting direction according to the cutting sequence number, and filling each adjacent two-dimensional segmentation image into a fixed thickness of volume data by using linear interpolation to obtain the segmentation volume data in the cutting direction;

[0018] fusing the segmentation volume data obtained in each cutting direction to obtain three-dimensional segmentation volume data.

[0019] Further, the segmentation volume data obtained in each direction is fused to obtain three-dimensional segmentation volume data by performing the following steps:

[0020] obtaining the values of each segmentation volume data at the same position voxel point;

[0021] selecting the two largest values in each voxel point value and calculating the average value, and taking the average value as the segmentation value of the corresponding position voxel point in the three-dimensional segmentation volume data, thereby obtaining the three-dimensional segmentation volume data.

[0022] Further, the target object in the three-dimensional segmentation volume data is filled by using the anchor box, and an anchor box marker is added to the original three-dimensional volume data, comprising:

[0023] taking the voxel point with the same segmentation value as any category value in the three-dimensional segmentation volume data as the effective voxel of the corresponding target object category;

[0024] selecting an anchor point with a spatial interval of m voxels in each effective voxel of the target object, and setting anchor boxes with different scales and aspect ratios based on the spatial interval of the anchor point, wherein m≥1;

[0025] selecting the anchor point as the anchor box vertex in the corresponding effective voxel based on the different scales and aspect ratios of the set anchor boxes, and generating different anchor boxes in descending order, thereby filling the effective voxel of the corresponding target object;

[0026] obtaining the edge position voxel point of each anchor box in the three-dimensional segmentation volume data, and adding an anchor box marker to the voxel point at the same position in the original three-dimensional volume data, wherein the anchor box marker is the category of the corresponding target object.

[0027] Further, by inputting the marked original three-dimensional volume data into the 3D CNN network, the voxel points around the edge are classified according to the pixel value and category of the voxel points inside the edge of the anchor box marker, thereby obtaining a three-dimensional volume data image with target object markers.

[0028] Further, it also includes cropping the marked original three-dimensional volume data to obtain each sub-volume data, inputting each sub-volume data into the 3D CNN network to obtain sub-volume data with target object markers, and further obtaining a three-dimensional volume data image with target object markers.

[0029] In another aspect, the embodiment of the present application provides a target object segmentation system for security CT three-dimensional images, comprising:

[0030] A data acquisition and processing module is configured to acquire original three-dimensional volume data of a security CT image, slice the original three-dimensional volume data to obtain a plurality of two-dimensional slice images;

[0031] A three-dimensional segmentation volume data acquisition module is configured to obtain each two-dimensional segmentation image based on a target object based on the plurality of two-dimensional slice images, and fuse each two-dimensional segmentation image to obtain three-dimensional segmentation volume data;

[0032] A data label adding module is configured to fill the target object in the three-dimensional segmentation volume data with an anchor box, and add an anchor box label to the original three-dimensional volume data;

[0033] A segmentation data acquisition module is configured to obtain a three-dimensional volume data image with a target object label based on the labeled original three-dimensional volume data.

[0034] Further, the data acquisition and processing module obtains the plurality of two-dimensional slice images by performing the following operations:

[0035] According to a three-dimensional Cartesian coordinate system, the original three-dimensional volume data is cut into slice volume data with a fixed thickness along the X, Y and Z directions respectively, and is sequentially numbered; the numbering includes a cutting direction and a cutting sequence number to which the slice volume data belongs;

[0036] Each slice volume data is projected along the thickness direction to obtain a plurality of two-dimensional slice images, wherein the pixel value of each pixel point of the two-dimensional slice image is the maximum value of the pixel values of the voxels at the corresponding position of the slice volume data in the thickness direction.

[0037] Compared with the prior art, the present application can achieve the following beneficial effects:

[0038] The target object segmentation method and system for security CT three-dimensional images provided by the present application can greatly reduce the calculation complexity, improve the segmentation accuracy and improve the calculation efficiency by slicing the original three-dimensional volume data of the security CT, converting the three-dimensional volume data into two-dimensional slice images, performing target object segmentation on the two-dimensional slice images and then fusing to obtain three-dimensional segmentation volume data; the target object is filled with an anchor box, the anchor box label is added to the original three-dimensional volume data, and then the original three-dimensional volume data is classified more finely, which greatly reduces the calculation amount of three-dimensional classification, further improves the segmentation accuracy, and enables the application in practice.

[0039] The technical solutions in the present application can be combined with each other to realize more preferred combination solutions. Other features and advantages of the present application will be described in the following description, and some advantages will become apparent from the description, or will be understood by those skilled in the art through implementation of the present application. The objects and other advantages of the present application can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0040] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated in and constitute a part of this application. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.

[0041] Figure 1 A flowchart of a target object segmentation method for a security CT three-dimensional image provided by an embodiment of the present application is shown in the figure.

[0042] Figure 2 A flowchart of original three-dimensional volume data segmentation provided by an embodiment of the present application is shown in the figure.

[0043] Figure 3 A structure diagram of a 2D UNet provided by an embodiment of the present application is shown in the figure.

[0044] Figure 4 A flowchart of obtaining three-dimensional segmentation volume data provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present application will be specifically described below with reference to the accompanying drawings, wherein the drawings constitute a part of this application and are used to explain the principles of the embodiments of the present application, but are not used to limit the scope of the present application.

[0046] Embodiment 1

[0047] One specific embodiment of the present application discloses a target object segmentation method for a security CT three-dimensional image, as shown in the figure, comprising the following steps: Figure 1 As shown in the figure, comprising the following steps:

[0048] S1, obtaining original three-dimensional volume data of a security CT image, slicing the original three-dimensional volume data to obtain a plurality of two-dimensional slice images.

[0049] Specifically, 3D luggage data generated in a security device is obtained as original three-dimensional volume data; it can be understood that the 3D data size collected by the security device CT of the same model is consistent.

[0050] In implementation, a plurality of two-dimensional slice images are obtained by performing the following operations:

[0051] According to a three-dimensional Cartesian coordinate system, the original three-dimensional volume data is cut into slice volume data with a fixed thickness along the X, Y and Z directions respectively, and sequentially numbered; the numbering includes the cutting direction and the cutting sequence number of the slice volume data;

[0052] Each slice volume data is projected along the thickness direction to obtain a plurality of two-dimensional slice images, wherein the pixel value of each pixel point of the two-dimensional slice image is the maximum value of the pixel value of each voxel at the corresponding position of the slice volume data in the thickness direction.

[0053] Specifically, the thickness of the slice volume data is n voxels, and the value of n is set according to the size of the original three-dimensional volume data obtained in practice; for example, the size of the original three-dimensional volume data is 750x750x750, and n is set to 10 voxels.

[0054] That is, the original three-dimensional volume data V obtained, wherein, N x , N y , N z is the number of voxels of the length, height and width of the original three-dimensional volume data, respectively, as shown in Figure 2 Taking the Z axis as the cutting direction, N z / n two-dimensional slice images parallel to the XY plane are generated: first, cutting with a fixed thickness of n voxels, N z / n Nx×Ny×n slice volume data are obtained, each slice volume data is projected along the thickness direction, and the maximum number operation is applied to each position in the thickness direction to obtain Nx×Ny two-dimensional slice images; in the same way, two-dimensional slice images in the X and Y directions are obtained, and three groups of two-dimensional slice images are obtained according to the three cutting directions N x / n+N y / n+N z / n two-dimensional slice images.

[0055] It can be understood that the extraction position of each two-dimensional slice image can be obtained according to the cutting sequence number of the slice volume data, that is, the depth information is retained, so that complete volume data can be accurately obtained when generating the cutting volume data subsequently.

[0056] S2, based on a plurality of two-dimensional slice images, obtain two-dimensional segmentation images based on a target object, and fuse each two-dimensional segmentation image to obtain three-dimensional segmentation volume data.

[0057] In implementation, multiple two-dimensional slice images are input into a 2D UNet network, and each two-dimensional slice image is segmented based on the target object category to obtain two-dimensional segmented images based on the target object. The value of each pixel in the two-dimensional segmented image is either a category value or a background value, where the category value represents the importance level of the target object. It should be noted that the more important the target object in the two-dimensional segmented image, the higher the category value. The importance level of the target object is set according to the actual application scenario. For example, if the application scenario is the identification of dangerous goods in security checks, then the importance level of the target object represents the hazard level of the dangerous goods.

[0058] Specifically, such as Figure 3 As shown, the 2D UNet network includes eight fully convolutional layers (conv1-conv8) as downsampling paths for feature extraction and eight deconvolutional layers (deconv1-deconv8) as upsampling paths for feature fusion. There is a skip connection between each convolutional and deconvolutional layer. In the downsampling paths, each convi (i = 1, ..., 8) layer outputs a spatial resolution 2π lower than the input image. i Feature map and 32×2 i The feature channels are defined such that each fully convolutional layer is followed by an activation layer, activated by the LeakyReLu activation function and batch normalized, with a bottleneck feature resolution of 1×1. In the upsampling path, the output of each deconvi (i = 8, ..., 1) layer is lower than the spatial resolution of the output image by 2. i-1 The features are such that it ultimately has the same number of channels as the number of object classes, wherein each deconvolution layer is followed by a ReLU activation layer and a batch normalization layer.

[0059] Specifically, before obtaining the two-dimensional segmented images based on the target object using the 2D UNet network, the 2D UNet network needs to be trained: two-dimensional slice images containing the target object and two-dimensional slice images not containing the target object are used as training samples and input into the 2D UNet network. The focus loss function is used as the training criterion to obtain the trained 2D UNet network. More specifically, two-dimensional slice images containing the target object in each bag are used as positive samples, and those not containing the target object are used as negative samples. The negative samples are obtained by randomly selecting several two-dimensional slice images from each bag; preferably, nine two-dimensional slice images are randomly selected from each bag. It should be noted that during training, two-dimensional slice images containing target objects of different importance levels are selected. The pixel values ​​in the resulting two-dimensional segmented images are either background values ​​or category values ​​of different importance levels.

[0060] When implementing, such as Figure 4As shown, each two-dimensional segmentation image is fused to obtain three-dimensional segmentation body data by performing the following steps:

[0061] The two-dimensional segmentation image in each segmentation direction is processed as follows to obtain the corresponding segmentation body data:

[0062] The two-dimensional segmentation image in the segmentation direction is arranged according to the segmentation sequence number, and each adjacent two-dimensional segmentation image is filled with a fixed thickness of body data using linear interpolation to obtain the segmentation body data in the segmentation direction;

[0063] The segmentation body data obtained in each segmentation direction is fused to obtain three-dimensional segmentation body data.

[0064] In specific implementation, the segmentation body data obtained in each direction is fused to obtain three-dimensional segmentation body data by performing the following steps:

[0065] The values of each segmentation body data at the same position voxel point are obtained;

[0066] The two largest values in each voxel point value are selected, and the average value is calculated as the segmentation value of the corresponding position voxel point of the three-dimensional segmentation body data, thereby obtaining the three-dimensional segmentation body data.

[0067] It can be understood that when obtaining the three-dimensional segmentation body data, the two largest values in each segmentation body data are selected for averaging, which ensures the spatial consistency between multiple segmentation directions and makes the 3D segmentation of each voxel more reliable. Moreover, since the image segmentation is performed using a 2D segmentation network and then fused, the computational complexity is reduced, the computational efficiency is improved, and the calculation time is shortened. In addition, the target object can be preliminarily segmented, so that subsequent labeling can be added to the original three-dimensional body data and sent to a classification network for further refinement, thereby improving the computational efficiency and segmentation accuracy.

[0068] S3, using an anchor box to fill the target object in the three-dimensional segmentation body data, and adding an anchor box label to the original three-dimensional body data.

[0069] In implementation, the use of an anchor box to fill the target object in the three-dimensional segmentation body data and adding an anchor box label to the original three-dimensional body data includes:

[0070] The voxel points in the three-dimensional segmentation body data with the same segmentation value as any category value are regarded as effective voxels of the corresponding target object category;

[0071] An anchor point with a spatial interval of m voxels is selected in the effective voxels of each target object, and anchor boxes with different scales and aspect ratios are set based on the spatial interval of the anchor point, where m≥1;

[0072] The anchor points are selected as the anchor box vertices in the corresponding effective voxels based on the different scales and aspect ratios of the set anchor boxes, and different anchor boxes are generated from large to small, and the effective voxels of the corresponding target objects are filled;

[0073] The edge position voxel points of each anchor box in the three-dimensional segmentation body data are acquired, and anchor box labels are added to the same position voxel points in the original three-dimensional body data, wherein the anchor box label is the category of the corresponding target object.

[0074] Specifically, the spatial interval m is set according to the size of the original three-dimensional body data, and the different scales and aspect ratios of the anchor boxes are set based on the spatial interval m according to experience, so as to meet the requirement of calculation amount.

[0075] For example, the size of the original three-dimensional body data is 750x750x750, the spatial interval m is set to 5, and the anchor box is set to 5 different scales and 8 aspect ratios, the scales are 5, 10, 15, 20 and 25 voxels, and the aspect ratios are 1:5, 2:5, 3:5, 4:5, 5:5, 2:3, 3:4 and 1:3.

[0076] S4, based on the labeled original three-dimensional body data, obtain a three-dimensional body data image with target object labels.

[0077] In implementation, the labeled original three-dimensional body data is input into the 3D CNN network, the voxel points around the edge are classified according to the pixel values and categories of the voxel points inside the anchor box label edge, and a three-dimensional body data image with target object labels is obtained. It can be understood that the 3D CNN network further classifies each pixel point of the original three-dimensional body data with anchor box labels, judges whether the voxel points around the anchor box label belong to the category according to the pixel values and categories of the voxel points surrounded by the anchor box label, refines the voxel points of the target object edge, and finally obtains a three-dimensional body data image in which the voxel points including the target object are all labeled with target object labels, and the target object segmentation of the original three-dimensional body data is completed.

[0078] Specifically, the 3D CNN network includes three convolutional blocks and two fully connected layers; wherein each convolutional block is composed of four 3D convolutional layers conv1-conv4, and a ReLU activation layer is set after each 3D convolutional layer, and the 3D convolutional layer conv1 of each convolutional block outputs a feature map with a spatial resolution lower than the input 3D feature map by 2 i and 64x2 i feature channels, and the conv2-conv4 of each convolutional block maintains the same resolution and channel number.

[0079] It should be noted that before obtaining the three-dimensional body data image with target object label through the 3D CNN network, training is needed to remove the areas only with cluttered background or with large overlap with the real target object.

[0080] Preferably, the method further comprises cropping the original three-dimensional body data after labeling to obtain each sub-body data, inputting each sub-body data into the 3D CNN network to obtain sub-body data with target object label, and further obtaining the three-dimensional body data image with target object label.

[0081] It can be understood that the original three-dimensional body data is cropped, and the size of each sub-body data after cropping is set according to the size of the original three-dimensional body data in the actual situation, so that the segmentation accuracy of the target object is met while greatly reducing the calculation amount in the 3D CNN network and speeding up the calculation speed.

[0082] Compared with the prior art, the target object segmentation method for the security CT three-dimensional image provided by the embodiment can greatly reduce the calculation complexity, improve the segmentation accuracy and improve the calculation efficiency by slicing the original three-dimensional body data of the security CT, converting the three-dimensional body data into two-dimensional slice images, performing target object segmentation on the two-dimensional slice images and then fusing to obtain three-dimensional segmentation body data; and by filling the target object with an anchor box and adding anchor box labels to the original three-dimensional body data, the original three-dimensional body data is classified more finely, the calculation amount of three-dimensional classification is greatly reduced, the segmentation accuracy is further improved, and the method can be applied in practice.

[0083] Embodiment 2

[0084] In one specific embodiment of the application, a target object segmentation system for security CT three-dimensional images is disclosed, comprising:

[0085] A data acquisition and processing module is configured to acquire original three-dimensional body data of a security CT image, slice the original three-dimensional body data to obtain a plurality of two-dimensional slice images.

[0086] A three-dimensional segmentation body data acquisition module is configured to obtain each two-dimensional segmentation image based on a target object based on the plurality of two-dimensional slice images, and fuse each two-dimensional segmentation image to obtain three-dimensional segmentation body data.

[0087] A data labeling adding module is configured to fill the target object in the three-dimensional segmentation body data with an anchor box and add anchor box labels to the original three-dimensional body data.

[0088] A segmentation data acquisition module is configured to obtain a three-dimensional body data image with target object label based on the original three-dimensional body data after labeling.

[0089] In implementation, the data acquisition and processing module obtains a plurality of two-dimensional slice images by performing the following operations:

[0090] According to a three-dimensional Cartesian coordinate system, the original three-dimensional body data is cut into slice body data of fixed thickness along X, Y and Z directions respectively, and is numbered in turn; the numbering comprises a cutting direction and a cutting sequence number to which the slice body data belongs;

[0091] Each slice body data is projected along the thickness direction to obtain a plurality of two-dimensional slice images, wherein a pixel value of each pixel point of the two-dimensional slice image is a maximum value of pixel values of voxels at a corresponding position of the slice body data in the thickness direction

[0092] The three-dimensional segmented body data acquisition module of the embodiment of the present application obtains three-dimensional segmented body data, and the specific implementation process of the data label adding module and the segmented data acquisition module can refer to the method embodiment described above, which will not be described here.

[0093] Since the embodiment has the same principle as the above-mentioned method embodiment, the system also has the corresponding technical effects of the above-mentioned method embodiment.

[0094] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory or a random access memory.

[0095] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A target object segmentation method for security CT three-dimensional images, characterized in that, The method comprises the following steps: Obtaining original three-dimensional body data of security CT images, slicing the original three-dimensional body data to obtain a plurality of two-dimensional slice images; Based on a plurality of two-dimensional slice images, obtaining each two-dimensional segmentation image based on a target object, and then fusing each two-dimensional segmentation image to obtain three-dimensional segmentation body data; Using an anchor box to fill the target object in the three-dimensional segmentation body data, and adding an anchor box mark to the original three-dimensional body data; comprising: The voxel points with the same segmentation value as any category value in the three-dimensional segmentation body data are regarded as effective voxels of the corresponding target object category; Selecting anchor points with a spatial interval of m voxels in the effective voxels of each target object, and setting anchor boxes with different scales and aspect ratios based on the spatial interval of the anchor points, wherein m≥1; Selecting the anchor points as the vertices of the anchor boxes in the corresponding effective voxels from large to small to generate different anchor boxes, and filling the effective voxels of the corresponding target object; Obtaining the edge position voxel points of each anchor box in the three-dimensional segmentation body data, and adding an anchor box mark to the voxel points at the same position in the original three-dimensional body data, wherein the anchor box mark is the corresponding target object category; Based on the marked original three-dimensional body data, a three-dimensional body data image with target object marks is obtained; wherein the marked original three-dimensional body data is input into a 3D CNN network, the voxel points around the edge are classified according to the pixel value and category of the voxel points inside the edge, and a three-dimensional body data image with target object marks is obtained; The 3D CNN network comprises three convolutional blocks and two fully connected layers; wherein each convolutional block is composed of four 3D convolutional layers conv1-conv4, a ReLU activation layer is arranged after each 3D convolutional layer, and the 3D convolutional layer conv1 of each convolutional block outputs a 3D feature map with a spatial resolution lower than the input by 2 i times and 64×2 i times the number of feature channels, wherein the conv2-conv4 of each convolutional block maintain the same resolution and the number of feature channels. 2.The method for target object segmentation of security CT three-dimensional images according to claim 1, characterized in that, A plurality of two-dimensional slice images are obtained by performing the following operations: According to a three-dimensional Cartesian coordinate system, the original three-dimensional body data is cut into slice body data with a fixed thickness along the X, Y and Z directions respectively, and is sequentially numbered; the numbering includes the cutting direction and the cutting sequence number of the slice body data; Projecting each slice body data along the thickness direction to obtain a plurality of two-dimensional slice images, wherein the pixel value of each pixel point of the two-dimensional slice image is the maximum pixel value of each voxel at the corresponding position of the slice body data in the thickness direction.

3. The method for target object segmentation of security CT three-dimensional images according to claim 1, characterized in that, By inputting a plurality of two-dimensional slice images into a 2D UNet network, each two-dimensional slice image is segmented based on the target object category to obtain each two-dimensional segmentation image based on the target object; wherein the value of each pixel point in the two-dimensional segmentation image is a category value or a background value, and the category value represents the importance level of the target object.

4. The target object segmentation method for security CT three-dimensional images according to claim 3, characterized in that, Each two-dimensional segmentation image is fused by performing the following steps to obtain three-dimensional segmentation body data: The two-dimensional segmentation images in each cutting direction are processed as follows to obtain the corresponding segmentation body data: Arrange the two-dimensional segmentation images in the cutting direction according to the cutting sequence number, and then fill each adjacent two-dimensional segmentation image into fixed-thickness body data using linear interpolation to obtain the segmentation body data in the cutting direction; Fuse the segmentation body data obtained in each cutting direction to obtain three-dimensional segmentation body data.

5. The method for target object segmentation of security CT three-dimensional images according to claim 4, characterized in that, Fuse the segmentation body data obtained in each direction to obtain three-dimensional segmentation body data by performing the following steps: Obtaining the value of each voxel point in the same position of the segmented data; Selecting the two largest values in each voxel point value and calculating the average value, which is taken as the segmentation value of the voxel point in the corresponding position of the three-dimensional segmented data, thereby obtaining the three-dimensional segmented data.

6. The method for target object segmentation of security CT three-dimensional images according to claim 1, characterized in that, Further comprising cropping the marked original three-dimensional data to obtain each sub-body data, inputting each sub-body data into the 3D CNN network to obtain sub-body data with target object labels, and further obtaining a three-dimensional data image with target object labels.

7. A target object segmentation system for security CT three-dimensional images, characterized by, The method comprises: a data acquisition and processing module configured to obtain original three-dimensional data of a security CT image, slice the original three-dimensional data to obtain a plurality of two-dimensional slice images, and obtain each two-dimensional segmentation image based on a target object based on the plurality of two-dimensional slice images; a three-dimensional segmented data acquisition module configured to fuse each two-dimensional segmentation image to obtain three-dimensional segmented data; a data label adding module configured to fill the target object in the three-dimensional segmented data with an anchor box and add an anchor box label to the original three-dimensional data; the method comprises: taking the voxel points with the same segmentation value as any category value in the three-dimensional segmented data as the effective voxels of the corresponding target object category; selecting anchor points with a spatial interval of m voxels in the effective voxels of each target object, and setting anchor boxes with different scales and aspect ratios based on the spatial interval of the anchor points, wherein m≥1; selecting the anchor points as the vertices of the anchor boxes in the corresponding effective voxels from large to small to generate different anchor boxes in sequence, and filling the effective voxels of the corresponding target object; obtaining the edge position voxel points of each anchor box in the three-dimensional segmented data, and adding an anchor box label to the voxel points in the same position in the original three-dimensional data, wherein the anchor box label is the corresponding target object category; a segmentation data acquisition module configured to obtain a three-dimensional data image with target object labels based on the marked original three-dimensional data; wherein the three-dimensional data image with target object labels is obtained by inputting the marked original three-dimensional data into a 3D CNN network and classifying the voxel points around the edge according to the pixel values and categories of the voxel points inside the edge. The 3D CNN network comprises three convolutional blocks and two fully connected layers; wherein each convolutional block is composed of four 3D convolutional layers conv1-conv4, a ReLU activation layer is arranged after each 3D convolutional layer, and the 3D convolutional layer conv1 of each convolutional block outputs a 3D feature map with a spatial resolution lower than the input by 2 i times and 64×2 i times the number of feature channels, wherein the conv2-conv4 of each convolutional block maintain the same resolution and the number of feature channels.

8. The target object segmentation system for security CT three-dimensional images of claim 7, wherein, The data acquisition and processing module obtains a plurality of two-dimensional slice images by performing the following operations: According to a three-dimensional Cartesian coordinate system, the original three-dimensional data is cut into slice body data with a fixed thickness along the X, Y, and Z directions, and is sequentially numbered; the numbering comprises the cutting direction and the cutting sequence number of the slice body data; projecting each slice body data along the thickness direction to obtain a plurality of two-dimensional slice images, wherein the pixel value of each pixel point of the two-dimensional slice image is the maximum value of the pixel values of the voxels in the corresponding position of the slice body data in the thickness direction.

Images