A method and system for extracting the outer contour of a three-dimensional security imaging object by sparse sampling

Abstract

This invention provides a method and system for extracting the outer contour of objects in sparsely sampled 3D security inspection imaging. The method includes: determining the Y-direction projection range and target projection surface based on the geometric relationship of the security CT scan and the position of the conveyor belt in the 3D volumetric data of the security CT scanner; constructing an XZ-plane super-resolution projection mesh; calculating the width and height of the projection mesh based on a preset super-resolution factor and the dimensional dimensions of the original 3D volumetric data; generating and initializing a corresponding 2D floating-point mesh as the accumulation volume for projection energy; performing XZ-plane Gaussian sputtering diffusion on the initial discrete energy map to obtain a normalized energy map; sequentially performing normalization, binarization segmentation, and morphological closing operations on the projection energy map; extracting the outermost contour of the image and filtering out noisy contours; mapping the retained contour coordinates back to the corresponding physical coordinates of the original 3D volumetric data; and outputting the outer contour of the object. This invention can significantly improve contour detection efficiency.

Description

Technical Field

[0001] This invention relates to the field of security inspection technology, and in particular to a method and system for extracting the outer contour of objects in sparse sampling three-dimensional security inspection imaging. Background Technology

[0002] Since its advent in the 1970s, computed tomography (CT) technology has expanded its applications from medical imaging diagnosis to industrial inspection, public safety, and other scenarios. With continuous advancements in detectors, X-ray sources, and image reconstruction algorithms, the speed, resolution, and imaging quality of CT scans have steadily improved. This has led to the development of security CT as an important branch in public safety fields such as airports, customs, and rail transportation. By performing 360° rotating scans on luggage and other items to acquire internal three-dimensional volumetric data, it provides core technological support for security inspections.

[0003] In the security inspection CT image processing workflow, the two-dimensional projection of three-dimensional volume data and the extraction of the object's outer contour are fundamental and crucial steps, providing a preliminary data foundation for subsequent advanced analyses such as item recognition, automatic detection of contraband, and alarms. Currently, three-dimensional volume data projection often employs weighted projection methods such as maximum density projection, average density projection, and direct slice projection; contour extraction is generally based on threshold segmentation of the projected image, combined with methods such as region growing, edge detection, or contour tracing.

[0004] However, existing technologies have significant limitations when applied to security CT scenarios: equal-weighted projection faces the sparse volume data in security scenarios, which are prone to problems such as broken and hollow projection images and jagged and discontinuous contours, and the extraction accuracy is limited by the original voxel spacing; some methods rely on point cloud clustering, shape prior matching and other processes, which have high computational complexity or are difficult to adapt to irregularly shaped objects, and existing solutions are mostly designed for medical CT, which are not adapted to the characteristics of sparse security CT data, complex object types and diverse shapes, and cannot meet the comprehensive requirements of security scenarios for contour extraction accuracy, completeness and real-time performance. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a method and system for extracting the outer contour of objects in sparse sampling three-dimensional security inspection imaging, so as to solve the problem that the existing technology is not adapted to the characteristics of sparse security inspection CT data, complex types of objects, and diverse shapes, which leads to the inability to meet the requirements of contour extraction accuracy, completeness and real-time performance in security inspection scenarios.

[0006] The first aspect of the present invention proposes: A method for extracting the outer contour of an object in sparsely sampled 3D security imaging, wherein the method includes: Acquire the three-dimensional volume data generated by the security inspection CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security inspection CT and the position of the conveyor belt; Construct an XZ plane super-resolution projection mesh. Based on the preset super-resolution factor and the dimensions of the original 3D volume data, calculate the width and height of the projection mesh, generate a 2D floating-point mesh of the corresponding size and complete the initialization, which serves as the accumulator of projection energy. Traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, weight the energy of the effective voxels and accumulate it to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. The projected energy map is then normalized, binarized, segmented, and morphologically closed sequentially. The outermost contour of the image is extracted and noise contours are filtered out. The remaining contour coordinates are mapped back to the physical coordinates corresponding to the original 3D volume data, and the outer contour of the object is output.

[0007] The beneficial effects of this invention are as follows: By combining the scanning geometry and the orientation of the conveyor belt to determine the projection range, an XZ-plane super-resolution projection mesh is constructed. Gaussian attenuation weighted accumulation of voxel energy, supplemented by planar Gaussian sputtering diffusion, effectively compensates for the sparseness of the original volume data, fills contour gaps, and improves contour continuity and integrity. The super-resolution mesh refines projection accuracy while controlling the computational scale, adapting to various complex security inspection items. Combined with normalization, binarization segmentation, and morphological closing operations, the outer contour is accurately extracted and noise is filtered out. Accurate three-dimensional physical contours are output through coordinate mapping, significantly improving the accuracy and integrity of object contour extraction in security inspection scenarios while ensuring real-time processing.

[0008] Furthermore, the step of constructing the XZ-plane super-resolution projection mesh specifically includes: The width dimension of the original 3D volume data corresponds to the width direction of the projection mesh, and the depth dimension of the original 3D volume data corresponds to the height direction of the projection mesh. The size of the projection mesh is scaled up proportionally to the size of the corresponding dimension of the original volume data according to the super-resolution factor. A two-dimensional floating-point grid of the corresponding size is generated, and the values ​​at all positions within the grid are initialized to zero to serve as the basis for subsequent projection energy accumulation.

[0009] Furthermore, the step of traversing voxels along the height direction and weighted accumulation specifically includes: For each position in the super-resolution projection grid, its width and depth coordinates in the original 3D volume data are obtained through coordinate mapping, and each voxel in the projection range is traversed along the height direction. Read the CT attenuation value of the voxel corresponding to the current layer, preset the effective voxel threshold, and determine the voxels with CT attenuation values ​​lower than the threshold as invalid voxels; Centered on the target projection plane, the height-direction attenuation weight corresponding to the current layer is calculated according to the Gaussian function law. The CT attenuation value of the effective voxel is multiplied by the corresponding weight and then accumulated to the corresponding position of the projection mesh.

[0010] Furthermore, before accumulating the voxel energy to the projected grid, a positive and negative energy field competition modulation step is included: The gradient magnitude in the height direction of the current voxel is calculated using the central difference method within the 3D volume data. At the boundary positions in the height direction, forward difference or backward difference is used instead of the central difference method. Pre-set object threshold and air threshold, and configure corresponding positive weight and negative weight. Calculate the modulated voxel energy based on the numerical relationship between the voxel CT attenuation value and the two thresholds. The modulated voxel energy is multiplied by the corresponding Gaussian decay weight in the height direction and then accumulated to the corresponding position of the projected grid to complete the competitive modulation process of the positive and negative energy fields.

[0011] Furthermore, the planar Gaussian sputtering diffusion process specifically includes: Set the Gaussian diffusion standard deviation of the XZ plane, and adjust its value according to the super-resolution factor and data sparsity. Control the diffusion intensity through preset coefficients to adapt to different data density application scenarios. Based on the standard deviation of Gaussian diffusion, a two-dimensional Gaussian convolution kernel with an odd size is determined. The two-dimensional Gaussian convolution operation is performed using a separable convolution method. First, a one-dimensional Gaussian convolution is performed along the width direction, and then a one-dimensional Gaussian convolution is performed along the depth direction. This reduces the amount of computation while achieving planar diffusion of energy.

[0012] Furthermore, the steps of the binarization segmentation and morphological closing operation specifically include: The diffused projected energy map is numerically linearly mapped and converted into an eight-bit unsigned integer image, so that the pixel values ​​are within the numerical range of the standard grayscale image. The eight-bit grayscale image is binarized using any one of the fixed threshold method, Otsu adaptive threshold method, or local mean adaptive threshold method to distinguish between the foreground object region and the background region. Morphological closing operation is then performed on the binarized image to fill in the tiny holes inside the image and optimize the integrity and continuity of the foreground object region.

[0013] Furthermore, the step of extracting the outer contour and filtering the noise contour specifically includes: Perform contour extraction on the morphologically processed binary image, set the retrieval mode to extract only the outermost contour, and set the contour approximation method to retain all pixels on the contour. The area, perimeter, and roundness of each extracted contour are calculated separately. Noisy contours with an area smaller than the minimum area threshold, a perimeter smaller than the minimum perimeter threshold, or a roundness lower than the preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted according to the super-resolution factor.

[0014] The second aspect of the present invention proposes: A sparse-sampling 3D security inspection imaging object contour extraction system, wherein the system comprises: The scanning module is used to acquire the three-dimensional volume data generated by the security CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security CT and the position of the conveyor belt. The construction module is used to construct the XZ plane super-resolution projection mesh. Based on the preset super-resolution multiple and the dimensions of the original 3D volume data, the width and height of the projection mesh are calculated, a 2D floating-point mesh of the corresponding size is generated and initialized, and it serves as the accumulator of projection energy. The calculation module is used to traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, accumulate the energy of the effective voxels in a weighted manner to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. The mapping module is used to perform normalization, binarization segmentation and morphological closing operation on the projected energy map in sequence, extract the outermost contour of the image and filter out the noise contour, map the retained contour coordinates back to the physical coordinates corresponding to the original 3D volume data, and output the outer contour of the object.

[0015] Furthermore, the step of constructing the XZ-plane super-resolution projection mesh specifically includes: The width dimension of the original 3D volume data corresponds to the width direction of the projection mesh, and the depth dimension of the original 3D volume data corresponds to the height direction of the projection mesh. The size of the projection mesh is scaled up proportionally to the size of the corresponding dimension of the original volume data according to the super-resolution factor. A two-dimensional floating-point grid of the corresponding size is generated, and the values ​​at all positions within the grid are initialized to zero to serve as the basis for subsequent accumulation of projected energy.

[0016] Furthermore, the step of traversing voxels along the height direction and weighted accumulation specifically includes: For each position in the super-resolution projection grid, its width and depth coordinates in the original 3D volume data are obtained through coordinate mapping, and each voxel in the projection range is traversed along the height direction. Read the CT attenuation value of the voxel corresponding to the current layer, pre-set the effective voxel threshold, and determine the voxels with CT attenuation values ​​lower than the threshold as invalid voxels; Centered on the target projection plane, the height-direction attenuation weight corresponding to the current layer is calculated according to the Gaussian function law. The CT attenuation value of the effective voxel is multiplied by the corresponding weight and then accumulated to the corresponding position of the projection mesh.

[0017] Furthermore, before accumulating the voxel energy to the projected grid, a positive and negative energy field competition modulation step is included: The gradient magnitude in the height direction of the current voxel is calculated using the central difference method within the 3D volume data. At the boundary positions in the height direction, forward difference or backward difference is used instead of the central difference method. Pre-set object threshold and air threshold, and configure corresponding positive weight and negative weight. Calculate the modulated voxel energy based on the numerical relationship between the voxel CT attenuation value and the two thresholds. The modulated voxel energy is multiplied by the corresponding Gaussian decay weight in the height direction and then accumulated to the corresponding position of the projected grid to complete the competitive modulation process of the positive and negative energy fields.

[0018] Furthermore, the planar Gaussian sputtering diffusion process specifically includes: Set the Gaussian diffusion standard deviation of the XZ plane, and adjust its value according to the super-resolution factor and data sparsity. Control the diffusion intensity through preset coefficients to adapt to different data density application scenarios. Based on the standard deviation of Gaussian diffusion, a two-dimensional Gaussian convolution kernel with an odd size is determined. The two-dimensional Gaussian convolution operation is performed using a separable convolution method. First, a one-dimensional Gaussian convolution is performed along the width direction, and then a one-dimensional Gaussian convolution is performed along the depth direction. This reduces the amount of computation while achieving planar diffusion of energy.

[0019] Furthermore, the steps of the binarization segmentation and morphological closing operation specifically include: The diffused projected energy map is numerically linearly mapped and converted into an eight-bit unsigned integer image, so that the pixel values ​​are within the numerical range of the standard grayscale image. The eight-bit grayscale image is binarized using any one of the fixed threshold method, Otsu adaptive threshold method, or local mean adaptive threshold method to distinguish between the foreground object region and the background region. Morphological closing operation is then performed on the binarized image to fill in the tiny holes inside the image and optimize the integrity and continuity of the foreground object region.

[0020] Furthermore, the step of extracting the outer contour and filtering the noise contour specifically includes: Perform contour extraction on the morphologically processed binary image, set the retrieval mode to extract only the outermost contour, and set the contour approximation method to retain all pixels on the contour. The area, perimeter, and roundness of each extracted contour are calculated separately. Noisy contours with an area smaller than the minimum area threshold, a perimeter smaller than the minimum perimeter threshold, or a roundness lower than the preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted according to the super-resolution factor.

[0021] The third aspect of the present invention proposes: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the sparse sampling three-dimensional security inspection imaging object contour extraction method as described above.

[0022] The fourth aspect of the present invention proposes: A readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the sparse sampling three-dimensional security inspection imaging object contour extraction method as described above.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] Figure 1 This is a flowchart of the sparse sampling three-dimensional security imaging object contour extraction method provided in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the sparse sampling three-dimensional security inspection imaging object contour extraction system provided in the third embodiment of the present invention.

[0025] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0026] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0027] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0029] Please see Figure 1 The figure shows the sparse sampling three-dimensional security inspection imaging object contour extraction method provided in the first embodiment of the present invention. The sparse sampling three-dimensional security inspection imaging object contour extraction method provided in this embodiment can accurately extract the outer contour and filter out noise by combining normalization, binarization segmentation and morphological closing operation. The accurate three-dimensional physical contour is output through coordinate mapping. Under the premise of ensuring real-time processing, the accuracy and completeness of object contour extraction in security inspection scenarios are significantly improved.

[0030] Specifically, this embodiment provides: A method for extracting the outer contour of an object in sparsely sampled 3D security imaging, wherein the method includes: Step S10: Obtain the three-dimensional volume data generated by the security CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security CT and the position of the conveyor belt. It's important to note that acquiring the 3D volumetric data generated by a security CT scan involves determining the projection range and target projection plane in the height direction based on the scanning geometry and conveyor belt position of the CT scanner. The 3D volumetric data from a security CT scan is composed of stacked tomographic images from consecutive scans. In sparse sampling mode, the interlayer spacing is much larger than in dense sampling, resulting in fewer voxels and more discrete data along the height direction. The projection range determines the voxel intervals involved in the projection calculation, avoiding computational interference from irrelevant areas. The target projection plane serves as the reference plane for the projection, typically chosen as the center plane of the object, ensuring the symmetry and clarity of the projected outline, maximizing the representation of the object's true dimensions, and avoiding dimensional distortion caused by forward or backward offsets.

[0031] Taking a typical subway baggage security CT scan as an example, the three-dimensional volume data includes three dimensions: X (horizontal width), Y (scanning depth, conveyor belt direction), and Z (vertical height). The resolution of a single frame tomographic image is 512×512 pixels (X×Z). There are 64 scan layers along the Y direction with a layer spacing of 5mm, which is typical sparse sampling data. The layer spacing is 10 times that of dense sampling, increasing the scanning speed by 4 times and reducing the radiation dose by 62%. The projection range covers the entire baggage area along the Y direction, covering a total of 48 effective voxels, with 8 layers reserved at the top and bottom to avoid contour truncation caused by object offset. The target projection surface is selected from the center layer in the Y direction, which is the center position of the baggage thickness, ensuring that the projected contour best represents the true outer dimensions of the object and avoiding dimensional distortion caused by forward and backward offset. The XZ plane is chosen as the projection plane because in the security scene, the conveyor belt moves along the Y direction, and the XZ plane is a cross-section perpendicular to the conveyor belt. The contour of this cross-section best reflects the cross-sectional shape of the object and is the core basis for identifying the shape and size of prohibited items. For example, the XZ cross-sectional contour features of prohibited items such as controlled knives, replica firearms, and flammable and explosive containers are the key basis for identification. Therefore, the output contour can be directly connected to the subsequent prohibited item matching algorithm, which meets the actual business needs of security inspection without the need for additional coordinate transformation.

[0032] Step S20: Construct an XZ plane super-resolution projection mesh. Calculate the width and height of the projection mesh based on the preset super-resolution factor and the dimensions of the original three-dimensional volume data. Generate a two-dimensional floating-point mesh of the corresponding size and complete its initialization, which serves as the accumulator of projection energy. It's important to note that the construction of the XZ-plane super-resolution projection mesh involves calculating the width and height of the projection mesh based on a preset super-resolution factor and the dimensions of the original 3D volume data. A corresponding 2D floating-point mesh is then generated and initialized, serving as the accumulation volume for projection energy. The super-resolution mesh is the core foundation for improving contour accuracy. By proportionally enlarging the projection dimension, the resolution is improved at the sub-pixel level, compensating for the insufficient accuracy caused by the sparsity of the original data and providing a higher-resolution base for subsequent smoothing processing. Compared to direct projection at the original resolution, super-resolution projection can improve the dimensional accuracy of the contour by 2-3 times, significantly reducing jagged edges while preserving all the information from the original data.

[0033] Step S30: Traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, accumulate the energy of the effective voxels in weighted manner to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. It's important to note that the process involves traversing each voxel layer within the projection range along the height direction. Using the target projection surface as the center, the weight of each voxel layer is calculated according to a Gaussian decay law. The energy of the effective voxels is then weighted and accumulated to the corresponding positions in the projection grid. Finally, a planar Gaussian sputtering diffusion process is performed on the accumulated discrete energy map to obtain a continuous projection energy map. The Gaussian weighting ensures that voxels closer to the target projection surface have higher weights and those farther away have lower weights, thus focusing on the contour information of the target layer and reducing contour blurring caused by overlapping objects. Planar Gaussian sputtering smooths the discrete voxel projection energy, filling the energy gaps caused by sparse sampling, making the projection energy map continuous and smooth, and eliminating jagged edges caused by discrete points. The name "sputtering" originates from the fact that this process resembles sputtering and diffusing discrete energy points in all directions to form a continuous energy distribution, distinguishing it from ordinary whole-image Gaussian smoothing.

[0034] Step S40: Normalize, binarize and morphologically close the projected energy map in sequence, extract the outermost contour of the image and filter out the noise contour, map the retained contour coordinates back to the physical coordinates corresponding to the original three-dimensional volume data, and output the outer contour of the object.

[0035] It should be noted that the projected energy map is sequentially normalized, binarized, segmented, and subjected to morphological closing operations. The outermost contour of the image is extracted, and noisy contours are filtered out. The retained contour coordinates are mapped back to the physical coordinates corresponding to the original 3D volume data, outputting the object's outer contour. In the post-processing stage, the floating-point energy map is converted into a standard binary image. Morphological operations are used to optimize the contour continuity, and noisy contours are removed through geometric feature filtering. Finally, accurate object outer contour coordinates are obtained, which can be directly used for subsequent security inspection applications such as size measurement, contour matching, and contraband identification.

[0036] The second embodiment of the present invention provides: Furthermore, the step of constructing the XZ-plane super-resolution projection mesh specifically includes: The width dimension of the original 3D volume data corresponds to the width direction of the projection mesh, and the depth dimension of the original 3D volume data corresponds to the height direction of the projection mesh. The size of the projection mesh is scaled up proportionally to the size of the corresponding dimension of the original volume data according to the super-resolution factor. A two-dimensional floating-point grid of the corresponding size is generated, and the values ​​at all positions within the grid are initialized to zero to serve as the basis for subsequent projection energy accumulation.

[0037] It's important to note that the width dimension of the original 3D volume data corresponds to the width direction of the projected mesh, and the depth dimension corresponds to the height direction of the projected mesh. This dimensional mapping ensures a one-to-one correspondence between the spatial coordinates of the projected mesh and the original 3D data. This allows for rapid positioning during subsequent voxel energy accumulation, preventing contour deformation caused by coordinate misalignment. Specifically, the width dimension corresponds to the horizontal direction of the security image, i.e., the X-axis, directly determining the horizontal dimensional accuracy of the contour; the depth dimension corresponds to the scan stacking direction, i.e., the Y-axis (the direction of the conveyor belt's movement), which is the core dimension of sparse sampling and a key direction for improving super-resolution.

[0038] In this example, the original 3D volume data has 512 voxels in the width dimension (X-axis), corresponding to the horizontal width direction of the projected mesh; 64 layers in the depth dimension (Y-axis, scan layer direction), corresponding to the vertical height direction of the projected mesh; and the height dimension (Z-axis) is the vertical direction within the tomographic image, participating in the energy calculation of each voxel layer. This dimensional mapping ensures that the projection result matches the geometric proportions of the original object, preventing stretching distortion. The aspect ratio of the projected contour differs from the aspect ratio of the actual cross-section of the object by less than 1%.

[0039] The dimensions of the projected grid are scaled up proportionally to the dimensions of the original volume data in the corresponding dimensions using a super-resolution factor. The super-resolution factor is the core parameter controlling projection accuracy; a higher factor results in higher accuracy but also greater computational load, requiring a trade-off between security inspection accuracy requirements and available computing power. In sparse sampling scenarios, the original resolution of the depth dimension is low, so the super-resolution improvement primarily targets the depth direction, while the width dimension can be scaled up proportionally.

[0040] The performance and efficiency comparisons at different super-resolution ratios are as follows: 2x Super-Resolution: Mesh size 1024×128 pixels, computational load is 4 times that of the original, outline size error is about 2.1mm, suitable for embedded devices with limited computing power; 4x Super Resolution: Mesh size 2048×256 pixels, computational load is 16 times that of the original, contour size error is about 0.9mm, achieving the best balance between accuracy and efficiency, making it the first choice for general scenarios; 8x Super Resolution: The grid size is 4096×512 pixels, the computational load is 64 times that of the original, and the contour size error is about 0.6mm. It is suitable for high-precision detection scenarios, but it has high requirements for hardware computing power.

[0041] In this example, the preset super-resolution multiplier is 4x. Therefore, the width of the projected grid is 512×4=2048 pixels and the height is 64×4=256 pixels. The total number of pixels increases from the original 512×64=32768 to 2048×256=524288, which is a 16x increase in resolution. This provides a sufficient pixel base for subsequent smoothing and effectively reduces the jaggedness of the outline.

[0042] A two-dimensional floating-point mesh of the corresponding size is generated, and the values ​​at all positions within the mesh are uniformly initialized to zero, serving as the basis for subsequent projection energy accumulation. The floating-point data format is used to preserve the precision of energy accumulation and avoid quantization errors caused by integer truncation; the all-zero initialization ensures that energy accumulation starts from the baseline, preventing interference from initial values. Data precision can be selected according to the hardware platform: embedded devices use float32 single-precision floating-point to improve computational speed while maintaining accuracy; desktop servers can use float64 double-precision to further improve computational precision. In this example, the float32 format is used to generate a 2048×256 single-precision floating-point mesh, with all pixels initially set to 0.0. Subsequently, the energy of each voxel is weighted and accumulated to its corresponding coordinate position, ultimately forming a discrete projection energy distribution, where the value of each pixel corresponds to the projection energy intensity at that location.

[0043] Furthermore, the step of traversing voxels along the height direction and weighted accumulation specifically includes: For each position in the super-resolution projection grid, its width and depth coordinates in the original 3D volume data are obtained through coordinate mapping, and each voxel in the projection range is traversed along the height direction. Read the CT attenuation value of the voxel corresponding to the current layer, preset the effective voxel threshold, and determine the voxels with CT attenuation values ​​lower than the threshold as invalid voxels; Centered on the target projection plane, the height-direction attenuation weight corresponding to the current layer is calculated according to the Gaussian function law. The CT attenuation value of the effective voxel is multiplied by the corresponding weight and then accumulated to the corresponding position of the projection mesh.

[0044] It's important to note that for each location in the super-resolution projection mesh, its corresponding width and depth coordinates in the original 3D volume data are obtained through coordinate mapping. This process is repeated for each voxel layer within the projection range along the height direction. Coordinate mapping serves as a bridge connecting the projection mesh and the original 3D data. Using bilinear interpolation, the sub-pixel coordinates of the projection mesh are mapped to the integer coordinates of the original voxels, ensuring the accuracy of energy accumulation. Because the projection mesh is super-resolution, one original voxel corresponds to multiple pixels in the projection mesh, and energy is distributed to corresponding locations, ensuring the continuity of energy distribution.

[0045] The projection range can be adaptively adjusted according to the size of the object being inspected: for small luggage (such as carry-on bags), the projection range is adjusted to 32 layers to reduce computation and improve processing speed; for large luggage (such as 28-inch suitcases), the projection range is expanded to 64 layers to ensure complete coverage of the object and avoid outline truncation. In this example, a standard carry-on suitcase is used as the test object, with a projection range covering 48 voxels, leaving 8 layers of margin at the top and bottom. This ensures complete projection of the object while avoiding too many invalid voxels participating in the calculation, balancing accuracy and efficiency.

[0046] The CT attenuation value of the corresponding voxel in the current layer is read, and an effective voxel threshold is preset. Voxels with CT attenuation values ​​lower than this threshold are identified as invalid voxels. The purpose of effective voxel screening is to remove low-CT-value voxels from the air background and retain only high-CT-value voxels belonging to the object for energy accumulation, reducing the interference of background noise on the projected contour. The effective voxel threshold is usually set according to the scan dose and object type of the security CT scanner, slightly lower than the CT attenuation value of common organic materials, ensuring that all object voxels are included.

[0047] The CT values ​​for common substances range as follows: air approximately -1000 HU, fat approximately -100 to 0 HU, soft tissue approximately 0 to 100 HU, and metals approximately 300 to 3000 HU. In this example, the effective voxel threshold is set to -200 HU. Voxels with CT attenuation values ​​higher than -200 HU are considered effective, while those lower are considered air background and do not participate in energy accumulation. This effectively filters out most background noise, improving the signal-to-noise ratio of the projection by approximately 3 times and significantly reducing background interference in the contour. The threshold can be adaptively adjusted for different detection objects: baggage security check threshold -200 HU, human body security check threshold -100 HU, and metal object detection can be increased to 0 HU, adapting to the needs of different scenarios.

[0048] Centered on the target projection plane, the height-direction attenuation weights for the current layer are calculated according to the Gaussian function. The CT attenuation values ​​of the effective voxels are multiplied by the corresponding weights and then accumulated at the corresponding positions in the projection mesh. The core function of the Gaussian attenuation weights is spatial weighting: voxels closer to the target projection plane have larger weights and contribute more to the contour; voxels farther away have lower weights, reducing interference from overlapping objects on the contour. Compared to traditional maximum intensity projection (MIP), Gaussian weighted projection can better highlight the contours of the target layer, reduce the overlap blurring between foreground and background, and simultaneously fuse information from multiple voxels to compensate for the lack of interlayer information in sparse sampling.

[0049] In this example, the standard deviation of the Gaussian function is set to 1.5 times the original interlayer spacing. The weight at the target projection plane is 1.0, the weight at a distance of 1 layer is 0.82, the weight at a distance of 2 layers is 0.45, and the weight at a distance of 3 layers is 0.19. Voxels beyond 3 layers have a weight below 0.1 and are ignored in the calculation, thus reducing the computational load while ensuring contour quality. Taking a certain voxel layer as an example, the average CT of this voxel layer is 220 HU. It is 1 layer away from the target projection plane and has a weight of 0.82. Therefore, the energy contributed by a single voxel is 220 × 0.82 = 180.4, which is accumulated to the corresponding position on the projection mesh. Through weighted accumulation, a discrete energy distribution corresponding to the object contour is formed on the projection mesh, with higher energy closer to the contour center and gradually decreasing energy at the edges.

[0050] Furthermore, before accumulating the voxel energy to the projected grid, a positive and negative energy field competition modulation step is included: The gradient magnitude in the height direction of the current voxel is calculated using the central difference method within the 3D volume data. At the boundary positions in the height direction, forward difference or backward difference is used instead of the central difference method. Pre-set object threshold and air threshold, and configure corresponding positive weight and negative weight. Calculate the modulated voxel energy based on the numerical relationship between the voxel CT attenuation value and the two thresholds. The modulated voxel energy is multiplied by the corresponding Gaussian decay weight in the height direction and then accumulated to the corresponding position of the projected grid to complete the competitive modulation process of the positive and negative energy fields.

[0051] It should be noted that the height gradient magnitude of the current voxel is calculated using the central difference method within the 3D volumetric data. However, at the height boundaries, forward or backward differencing is used instead of the central difference method. The gradient magnitude reflects the degree of change in CT value along the height direction; the gradient is large at the object's boundaries and small inside the object compared to the background. The central difference method calculates the gradient by subtracting the values ​​of two layers, offering higher accuracy. However, since there are no layers at the boundaries, forward or backward differencing is used instead to ensure accurate gradient calculation at these boundaries.

[0052] In this example, the formula for calculating the height gradient is: internal voxel gradient = (voxel(i+1) - voxel(i-1)) / 2, in HU / layer; top boundary voxel gradient = voxel(i+1) - voxel(i); bottom boundary voxel gradient = voxel(i) - voxel(i-1). The gradient inside the object is usually less than 50 HU / layer, while the gradient at the boundary can reach 150~200 HU / layer, showing a significant difference. The boundary position can be accurately located using the gradient.

[0053] Preset an object threshold and an air threshold, and configure corresponding positive weights and negative weights at the same time. The modulated voxel energy is calculated according to the numerical relationship between the voxel CT attenuation value and the two thresholds. The core principle of dual-threshold competitive modulation is: voxels with CT values higher than the object threshold belong to the interior of the object, are assigned positive weights, and accumulate positive energy; voxels with CT values lower than the air threshold belong to the background air, are assigned negative weights, and accumulate negative energy; the area between the two thresholds is the boundary transition zone, which is weighted in combination with the gradient value. The larger the gradient, the closer to the boundary, and the higher the weight. Through the competition of positive and negative energy, the positive energy inside the object is superimposed, the negative energy of the background is counteracted, and a steep energy transition is formed at the edge, which greatly improves the contrast of the contour, and the edge definition is increased by more than 40% compared with simple effective voxel screening.

[0054] In this example, the object threshold is set to 100HU, corresponding to a positive weight of 1.2; the air threshold is set to -500HU, corresponding to a negative weight of -1.0. The modulation rules are: CT value ≥ object threshold: modulated energy = CT value × positive weight; CT value ≤ air threshold: modulated energy = CT value × negative weight (the result is a negative value); Air threshold < CT value < object threshold: transition zone weight = (CT value - air threshold) / (object threshold - air threshold) × gradient normalization coefficient, modulated energy = CT value × transition zone weight.

[0055] Taking the test sample as an example, before modulation, the average energy inside the object in the projection energy map is 62, the average energy in the background area is 21, and the contrast is about 2.95:1; after modulation through positive and negative energy competition, the average energy inside the object is increased to 78, the average energy in the background area is reduced to -11, the contrast is increased to 7.09:1, the energy gradient in the edge area is increased by 1.2 times, the sharpness of the contour is greatly improved, and the mis-segmentation rate of subsequent binarization is reduced by 45%.

[0056] Multiply the modulated voxel energy by the corresponding Gaussian attenuation weight in the height direction, and then accumulate it to the corresponding position of the projection grid to complete the competitive modulation process of the positive and negative energy fields. Compared with simple positive energy accumulation, in the projection energy map after competitive modulation, the background noise is counteracted by the negative energy, the contrast between the object and the background is increased by more than 2 times, the edge gradient is steeper, which provides a better basis for subsequent segmentation.

[0057] Further, the step of the planar Gaussian sputtering diffusion processing specifically includes: Set the Gaussian diffusion standard deviation of the XZ plane, whose value is adjusted according to the super-resolution multiple and data sparsity, and the diffusion intensity is controlled by a preset coefficient to adapt to application scenarios of different data densities; Based on the standard deviation of Gaussian diffusion, a two-dimensional Gaussian convolution kernel with an odd size is determined. The two-dimensional Gaussian convolution operation is performed using a separable convolution method. First, a one-dimensional Gaussian convolution is performed along the width direction, and then a one-dimensional Gaussian convolution is performed along the depth direction. This reduces the amount of computation while achieving planar diffusion of energy.

[0058] It's important to note that the Gaussian diffusion standard deviation for the XZ plane is adjusted based on the super-resolution factor and data sparsity. A preset coefficient controls the diffusion intensity, adapting to different data density application scenarios. The Gaussian diffusion standard deviation is the core parameter controlling the degree of diffusion: a larger standard deviation results in stronger diffusion and a smoother contour, but with greater loss of detail; a smaller standard deviation results in weaker diffusion and a sharper contour, but with more pronounced dispersion. Higher sparsity and a larger super-resolution factor require a greater diffusion intensity, necessitating a higher standard deviation value.

[0059] Reference for selecting standard deviation under different sparsity levels: 32-layer ultra-sparse data: standard deviation of 3.5 pixels, ensuring sufficient connectivity of discrete points; 64 layers of sparse data (in this example): standard deviation of 2.5 pixels, balancing smoothness and detail; 128 layers of semi-sparse data: standard deviation of 1.5 pixels, preserving more details.

[0060] In this example, the super-resolution factor is 4, the original data has 64 layers, and the standard deviation of Gaussian diffusion is set to 2.5 pixels, which corresponds to approximately 0.625 layers of the original data. This effectively fills the gaps between discrete voxels without excessive blurring that would cause the outlines to become thicker. The core difference between "Gaussian sputtering diffusion" and ordinary Gaussian smoothing is that ordinary Gaussian smoothing is a noise reduction process for the entire continuous image, while this method first "sputters" the discrete voxel energy points, allowing the discrete energy points to connect with each other to form continuous energy regions. It is specifically designed for the discrete energy distribution after sparse projection, making it more targeted.

[0061] A two-dimensional Gaussian convolution kernel with an odd-numbered size is determined based on the Gaussian diffusion standard deviation. A separable convolution method is used to perform the two-dimensional Gaussian convolution operation: first, a one-dimensional Gaussian convolution is performed along the width direction, and then another one-dimensional Gaussian convolution is performed along the depth direction. This reduces computational cost while achieving planar energy diffusion. The separability of the two-dimensional Gaussian convolution is because the two-dimensional Gaussian function can be decomposed into the product of two one-dimensional Gaussian functions in the X and Z directions. Therefore, the two-dimensional convolution can be split into two one-dimensional convolutions, reducing the computational cost from O(n²) to O(n), significantly improving the computational speed and making it suitable for the real-time requirements of security inspection scenarios.

[0062] Taking a 2048×256 projected grid as an example, using a 5×5 Gaussian kernel direct 2D convolution requires 25 multiply-accumulate operations per pixel, totaling approximately 13.1 million operations for the entire image (2048×256×25). Using separable convolution, first performing a 1×5 convolution in the horizontal direction (5 operations per pixel), then a 5×1 convolution in the vertical direction (5 operations per pixel), the total computational cost is reduced to approximately 5.24 million operations (2048×256×(5+5)). On an embedded ARM platform, the single-frame processing time was reduced from 31ms to 18ms, fully meeting the real-time processing requirements of security inspection equipment (more than 3 frames per second).

[0063] After Gaussian sputtering diffusion, the originally discrete energy points are diffused into a continuous energy region, filling the "holes" caused by sparse sampling, effectively smoothing the jagged edges of the contour, and preserving the overall shape and dimensional accuracy of the contour. Compared with the discrete energy map before diffusion, the continuity of the image after diffusion is improved by more than 60%, the jagged edges of the contour are basically eliminated, and no obvious step-like faults are visible to the naked eye.

[0064] Furthermore, the steps of the binarization segmentation and morphological closing operation specifically include: The diffused projected energy map is numerically linearly mapped and converted into an eight-bit unsigned integer image, so that the pixel values ​​are within the numerical range of the standard grayscale image. The eight-bit grayscale image is binarized using any one of the fixed threshold method, Otsu adaptive threshold method, or local mean adaptive threshold method to distinguish between the foreground object region and the background region. Morphological closing operation is then performed on the binarized image to fill in the tiny holes inside the image and optimize the integrity and continuity of the foreground object region.

[0065] It's important to note that the diffused projection energy map undergoes a linear numerical mapping to convert it into an 8-bit unsigned integer image, ensuring that pixel values ​​fall within the range of standard grayscale images. Normalization employs a linear mapping method: first, the minimum and maximum values ​​of the energy map are calculated, mapping the minimum to 0 and the maximum to 255. Intermediate values ​​are then linearly interpolated to convert to an 8-bit grayscale image ranging from 0 to 255. After conversion to a standard grayscale image, mature digital image processing algorithms can be used for subsequent processing, offering greater compatibility while reducing data volume and improving processing speed. In this example, the minimum value of the projection energy map is -128, and the maximum value is 395, linearly mapped to 0-255. Here, 0 corresponds to a pure background, and 255 corresponds to the densest areas of objects, conforming to the numerical conventions of conventional grayscale images and facilitating subsequent algorithm calls.

[0066] Binarization segmentation is performed on an 8-bit grayscale image using any one of the following methods: fixed thresholding, Otsu adaptive thresholding, or local mean adaptive thresholding, to distinguish foreground object regions from background regions. Binarization is a crucial step in dividing a grayscale image into foreground (object) and background, and the choice of threshold directly affects the segmentation accuracy.

[0067] Comparison of applicable scenarios for the three binarization methods: Fixed threshold method: Pre-set a fixed threshold. It has the fastest processing speed, about 2ms faster than the Otsu method. It is suitable for batch detection scenarios of the same batch and type of luggage, such as the same size packages in logistics security inspection. However, it has poor adaptability and large errors when the brightness changes. Otsu's adaptive thresholding method: It automatically calculates the threshold that maximizes the inter-class variance, resulting in high overall segmentation accuracy and strong adaptability. It is the preferred method for general scenarios. In this example, this method is used, and the segmentation accuracy is 15.2% higher than that of a fixed threshold. Local mean adaptive thresholding method: It uses the mean of the local neighborhood as the threshold, which is suitable for scenarios with uneven brightness, such as large luggage, human security checks, and situations where the edges are darker and the center is brighter. However, it requires more computation and is about 5ms slower than Otsu's method.

[0068] In this example, the Otsu method is used to automatically calculate the segmentation threshold, which is set to 128. Pixels above the threshold are identified as foreground objects, while those below are identified as background, resulting in a binarized image. Compared to a fixed threshold, the Otsu method improves segmentation accuracy by approximately 15%, effectively avoiding oversegmentation or undersegmentation, and preventing issues such as missing contours or excessive background noise.

[0069] Morphological closing operations are performed on the binarized image to fill tiny holes within the image, optimizing the integrity and continuity of foreground object regions. The morphological closing operation first dilates and then erodes, filling tiny holes smaller than the structuring element and connecting adjacent broken areas, making the foreground object's contour more continuous and its interior more complete, thus resolving the contour breaks and internal voids caused by sparse projection. The size of the structuring element is adjusted according to the super-resolution factor to ensure appropriate filling effects without excessive contour dilation.

[0070] A 3×3 square structuring element was used to perform the closing operation. Before processing, the binarized object contour had four small breaks with a spacing of 2-3 pixels, and seven holes with a diameter of 1-3 pixels inside. After the closing operation, all breaks were connected, the hole filling rate reached over 85%, the continuity of the object region was improved by 32%, and the integrity of the contour was significantly improved, providing a high-quality binary image for subsequent outer contour extraction. For sparser data, a 5×5 structuring element can be used to ensure the connection effect of broken contours.

[0071] Furthermore, the step of extracting the outer contour and filtering the noise contour specifically includes: Perform contour extraction on the morphologically processed binary image, set the retrieval mode to extract only the outermost contour, and set the contour approximation method to retain all pixels on the contour. The area, perimeter, and roundness of each extracted contour are calculated separately. Noisy contours with an area smaller than the minimum area threshold, a perimeter smaller than the minimum perimeter threshold, or a roundness lower than the preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted according to the super-resolution factor.

[0072] It's important to note that contour extraction is performed on the morphologically processed binary image. The retrieval mode is set to extract only the outermost contour, and the contour approximation method is set to retain all pixels on the contour. Extracting only the outermost contour avoids interference from internal contours caused by holes and textures within the object, retaining only the object's outer boundary contour, which meets the requirements of security inspection scenarios for the object's outer contour. Retaining all contour pixels ensures the accuracy of the contour and prevents the loss of details due to approximation. This example uses the Suzuki contour tracking algorithm, which can extract all contours in a single traversal, making it fast and suitable for real-time processing. A closed sequence of outer contour pixels, containing approximately 1800 pixels, is extracted, completely surrounding the object region. Internal hole contours are completely excluded, avoiding internal interference, and the single-pixel accuracy of the contour ensures the accuracy of subsequent size calculations.

[0073] For each extracted contour, the area, perimeter, and roundness are calculated separately. Noisy contours with areas smaller than a minimum area threshold, perimeters smaller than a minimum perimeter threshold, or roundness lower than a preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted based on the super-resolution factor. Geometric feature filtering is key to removing noisy contours: after sparse sampling and diffusion processing, small noise points may remain in the image, forming pseudo-contours. By filtering based on area, perimeter, and roundness, these noises can be effectively eliminated, retaining only the true object outline.

[0074] The minimum area threshold is adaptively adjusted by the square of the super-resolution factor, since area is the square of length. In this example, the super-resolution factor is 4, and the area is magnified 16 times. Therefore, the minimum area threshold is set to 200 pixels, corresponding to the original physical size of 12.5 pixels, which ensures that small noise is removed without missing small prohibited items. The minimum perimeter threshold is set to 60 pixels, and the roundness threshold is set to 0.2. The roundness calculation formula is C=4πS / L², and the closer it is to 1, the closer the outline is to a circle.

[0075] Besides filtering noise, geometric features can also be used for preliminary object classification: roundness greater than 0.7 indicates liquid containers or spherical objects; roundness between 0.4 and 0.7 indicates bags or suitcases; and roundness less than 0.3 indicates knives or other long, narrow tools. Preliminary roundness screening can quickly locate suspected knives and other prohibited items, improving subsequent identification efficiency. In this example, the roundness of the carry-on bag's outline is 0.72, which falls within the roundness range for bags or suitcases. If a long, narrow outline with a roundness of 0.28 is detected, it is automatically marked as a suspected knife and proceeds to subsequent fine-grained identification.

[0076] In this example, four contours were initially extracted. The object contour has an area of ​​approximately 260,000 pixels, a perimeter of 2,100 pixels, and a roundness of 0.72, which matches the characteristics of luggage contours. The remaining three smaller contours have areas of less than 150 pixels and roundness of less than 0.15, and were all filtered out. The noise contour removal rate reached over 90%.

[0077] The preserved contour coordinates are mapped back to the physical coordinates corresponding to the original 3D volume data, outputting the object's outer contour. The mapping process considers the super-resolution factor and the physical size of the original voxels, converting pixel coordinates into actual physical size coordinates, which can be directly used for subsequent applications such as size measurement and contraband contour matching.

[0078] In this example, the physical dimensions of the original 3D voxels are: 0.5 mm / pixel in the X direction, 0.5 mm / pixel in the Z direction, and 5 mm / layer in the Y direction (scanning layer direction). With a super-resolution factor of 4, the physical dimensions of the projected mesh are: 0.5 / 4 = 0.125 mm / pixel in the X direction and 5 / 4 = 1.25 mm / pixel in the Y direction. The maximum width of the extracted contour is 420 pixels, corresponding to a physical width of 420 × 0.125 = 52.5 mm; the maximum height is 180 pixels, corresponding to a physical height of 180 × 1.25 = 225 mm. Comparison with the actual object dimensions shows a measurement error of 0.87 mm, far superior to the 4.2 mm error of traditional methods, fully meeting the ±1 mm accuracy requirement for security inspection size detection. The output contour data can be directly interfaced with a security inspection intelligent recognition system for contour matching and size screening of contraband.

[0079] Please see Figure 2 The third embodiment of the present invention provides: A sparse-sampling 3D security inspection imaging object contour extraction system, wherein the system comprises: The scanning module is used to acquire the three-dimensional volume data generated by the security CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security CT and the position of the conveyor belt. The construction module is used to construct the XZ plane super-resolution projection mesh. Based on the preset super-resolution multiple and the dimensions of the original 3D volume data, the width and height of the projection mesh are calculated, a 2D floating-point mesh of the corresponding size is generated and initialized, and it serves as the accumulator of projection energy. The calculation module is used to traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, accumulate the energy of the effective voxels in a weighted manner to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. The mapping module is used to perform normalization, binarization segmentation and morphological closing operation on the projected energy map in sequence, extract the outermost contour of the image and filter out the noise contour, map the retained contour coordinates back to the physical coordinates corresponding to the original 3D volume data, and output the outer contour of the object.

[0080] Furthermore, the step of constructing the XZ-plane super-resolution projection mesh specifically includes: The width dimension of the original 3D volume data corresponds to the width direction of the projection mesh, and the depth dimension of the original 3D volume data corresponds to the height direction of the projection mesh. The size of the projection mesh is scaled up proportionally to the size of the corresponding dimension of the original volume data according to the super-resolution factor. A two-dimensional floating-point grid of the corresponding size is generated, and the values ​​at all positions within the grid are initialized to zero to serve as the basis for subsequent projection energy accumulation.

[0081] Furthermore, the step of traversing voxels along the height direction and weighted accumulation specifically includes: For each position in the super-resolution projection grid, its width and depth coordinates in the original 3D volume data are obtained through coordinate mapping, and each voxel in the projection range is traversed along the height direction. Read the CT attenuation value of the voxel corresponding to the current layer, preset the effective voxel threshold, and determine the voxels with CT attenuation values ​​lower than the threshold as invalid voxels; Centered on the target projection plane, the height-direction attenuation weight corresponding to the current layer is calculated according to the Gaussian function law. The CT attenuation value of the effective voxel is multiplied by the corresponding weight and then accumulated to the corresponding position of the projection mesh.

[0082] Furthermore, before accumulating the voxel energy to the projected grid, a positive and negative energy field competition modulation step is included: The gradient magnitude in the height direction of the current voxel is calculated using the central difference method within the 3D volume data. At the boundary positions in the height direction, forward difference or backward difference is used instead of the central difference method. Pre-set object threshold and air threshold, and configure corresponding positive weight and negative weight. Calculate the modulated voxel energy based on the numerical relationship between the voxel CT attenuation value and the two thresholds. The modulated voxel energy is multiplied by the corresponding Gaussian decay weight in the height direction and then accumulated to the corresponding position of the projected grid to complete the competitive modulation process of the positive and negative energy fields.

[0083] Furthermore, the planar Gaussian sputtering diffusion process specifically includes the following steps: Set the Gaussian diffusion standard deviation of the XZ plane, and adjust its value according to the super-resolution factor and data sparsity. Control the diffusion intensity through preset coefficients to adapt to different data density application scenarios. Based on the standard deviation of Gaussian diffusion, a two-dimensional Gaussian convolution kernel with an odd size is determined. The two-dimensional Gaussian convolution operation is performed using a separable convolution method. First, a one-dimensional Gaussian convolution is performed along the width direction, and then a one-dimensional Gaussian convolution is performed along the depth direction. This reduces the amount of computation while achieving planar diffusion of energy.

[0084] Furthermore, the steps of the binarization segmentation and morphological closing operation specifically include: The diffused projected energy map is numerically linearly mapped and converted into an eight-bit unsigned integer image, so that the pixel values ​​are within the numerical range of the standard grayscale image. The eight-bit grayscale image is binarized using any one of the fixed threshold method, Otsu adaptive threshold method, or local mean adaptive threshold method to distinguish between the foreground object region and the background region. Morphological closing operation is then performed on the binarized image to fill in the tiny holes inside the image and optimize the integrity and continuity of the foreground object region.

[0085] Furthermore, the step of extracting the outer contour and filtering the noise contour specifically includes: Perform contour extraction on the morphologically processed binary image, set the retrieval mode to extract only the outermost contour, and set the contour approximation method to retain all pixels on the contour. The area, perimeter, and roundness of each extracted contour are calculated separately. Noisy contours with an area smaller than the minimum area threshold, a perimeter smaller than the minimum perimeter threshold, or a roundness lower than the preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted according to the super-resolution factor.

[0086] The fourth embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the sparse sampling three-dimensional security inspection imaging object contour extraction method as described above.

[0087] The fifth embodiment of the present invention provides a readable storage medium storing a computer program thereon, wherein the program, when executed by a processor, implements the sparse sampling three-dimensional security inspection imaging object contour extraction method as described above.

[0088] In summary, the sparse sampling three-dimensional security inspection imaging object contour extraction method and system provided by the above embodiments of the present invention can accurately extract the outer contour and filter out noise by combining normalization, binarization segmentation and morphological closing operation, and output accurate three-dimensional physical contours through coordinate mapping. Under the premise of ensuring real-time processing, it significantly improves the accuracy and completeness of object contour extraction in security inspection scenarios.

[0089] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0090] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0091] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0092] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0093] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for extracting the outer contour of an object in sparse sampling three-dimensional security imaging, characterized in that, The method includes: Acquire the three-dimensional volume data generated by the security inspection CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security inspection CT and the position of the conveyor belt; Construct an XZ plane super-resolution projection mesh. Based on the preset super-resolution factor and the dimensions of the original 3D volume data, calculate the width and height of the projection mesh, generate a 2D floating-point mesh of the corresponding size and complete the initialization, which serves as the accumulator of projection energy. Traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, weight the energy of the effective voxels and accumulate it to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. The projected energy map is then normalized, binarized, segmented, and morphologically closed sequentially. The outermost contour of the image is extracted and noise contours are filtered out. The remaining contour coordinates are mapped back to the physical coordinates corresponding to the original 3D volume data, and the outer contour of the object is output.

2. The sparse sampling method for extracting the outer contour of an object in three-dimensional security imaging according to claim 1, characterized in that, The steps for constructing the XZ-plane super-resolution projection mesh specifically include: The width dimension of the original 3D volume data corresponds to the width direction of the projection mesh, and the depth dimension of the original 3D volume data corresponds to the height direction of the projection mesh. The size of the projection mesh is scaled up proportionally to the size of the corresponding dimension of the original volume data according to the super-resolution factor. A two-dimensional floating-point grid of the corresponding size is generated, and the values ​​at all positions within the grid are initialized to zero to serve as the basis for subsequent projection energy accumulation.

3. The sparse sampling method for extracting the outer contour of objects in three-dimensional security imaging according to claim 1, characterized in that, The step of traversing voxels along the height direction and weighted accumulating them specifically includes: For each position in the super-resolution projection grid, its width and depth coordinates in the original 3D volume data are obtained through coordinate mapping, and each voxel in the projection range is traversed along the height direction. Read the CT attenuation value of the voxel corresponding to the current layer, preset the effective voxel threshold, and determine the voxels with CT attenuation values ​​lower than the threshold as invalid voxels; Centered on the target projection plane, the height-direction attenuation weight corresponding to the current layer is calculated according to the Gaussian function law. The CT attenuation value of the effective voxel is multiplied by the corresponding weight and then accumulated to the corresponding position of the projection mesh.

4. The sparse sampling method for extracting the outer contour of objects in three-dimensional security imaging according to claim 1, characterized in that, Before the voxel energy is accumulated into the projected mesh, a positive and negative energy field competition modulation step is also included: The gradient magnitude in the height direction of the current voxel is calculated using the central difference method within the 3D volume data. At the boundary positions in the height direction, forward difference or backward difference is used instead of the central difference method. Pre-set object threshold and air threshold, and configure corresponding positive weight and negative weight. Calculate the modulated voxel energy based on the numerical relationship between the voxel CT attenuation value and the two thresholds. The modulated voxel energy is multiplied by the corresponding Gaussian decay weight in the height direction and then accumulated to the corresponding position of the projected grid to complete the competitive modulation process of the positive and negative energy fields.

5. The sparse sampling method for extracting the outer contour of an object in three-dimensional security imaging according to claim 1, characterized in that, The specific steps of the planar Gaussian sputtering diffusion process include: Set the Gaussian diffusion standard deviation of the XZ plane, and adjust its value according to the super-resolution factor and data sparsity. Control the diffusion intensity through preset coefficients to adapt to different data density application scenarios. Based on the standard deviation of Gaussian diffusion, a two-dimensional Gaussian convolution kernel with an odd size is determined. The two-dimensional Gaussian convolution operation is performed using a separable convolution method. First, a one-dimensional Gaussian convolution is performed along the width direction, and then a one-dimensional Gaussian convolution is performed along the depth direction. This reduces the amount of computation while achieving planar diffusion of energy.

6. The sparse sampling method for extracting the outer contour of an object in three-dimensional security imaging according to claim 1, characterized in that, The steps of the binarization segmentation and morphological closing operation specifically include: The diffused projected energy map is numerically linearly mapped and converted into an eight-bit unsigned integer image, so that the pixel values ​​are within the numerical range of the standard grayscale image. The eight-bit grayscale image is binarized using any one of the fixed threshold method, Otsu adaptive threshold method, or local mean adaptive threshold method to distinguish between the foreground object region and the background region. Morphological closing operation is then performed on the binarized image to fill in the tiny holes inside the image and optimize the integrity and continuity of the foreground object region.

7. The sparse sampling method for extracting the outer contour of an object in three-dimensional security imaging according to claim 1, characterized in that, The steps of extracting the outer contour and filtering the noise contour specifically include: Perform contour extraction on the morphologically processed binary image, set the retrieval mode to extract only the outermost contour, and set the contour approximation method to retain all pixels on the contour. The area, perimeter, and roundness of each extracted contour are calculated separately. Noisy contours with an area smaller than the minimum area threshold, a perimeter smaller than the minimum perimeter threshold, or a roundness lower than the preset roundness threshold are filtered out. The minimum area threshold is adaptively adjusted according to the super-resolution factor.

8. A sparse sampling three-dimensional security inspection imaging object contour extraction system, characterized in that, The system includes: The scanning module is used to acquire the three-dimensional volume data generated by the security CT scan, and determine the projection range and target projection surface in the height direction based on the scanning geometry of the security CT and the position of the conveyor belt. The construction module is used to construct the XZ plane super-resolution projection mesh. Based on the preset super-resolution multiple and the dimensions of the original 3D volume data, the width and height of the projection mesh are calculated, a 2D floating-point mesh of the corresponding size is generated and initialized, and it serves as the accumulator of projection energy. The calculation module is used to traverse each voxel layer within the projection range along the height direction, calculate the weight corresponding to each voxel layer according to the Gaussian decay law with the target projection surface as the center, accumulate the energy of the effective voxels in a weighted manner to the corresponding position of the projection grid, and then perform planar Gaussian sputtering diffusion processing on the accumulated discrete energy map to obtain a continuous projection energy map. The mapping module is used to perform normalization, binarization segmentation and morphological closing operation on the projected energy map in sequence, extract the outermost contour of the image and filter out the noise contour, map the retained contour coordinates back to the physical coordinates corresponding to the original 3D volume data, and output the outer contour of the object.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the sparse sampling three-dimensional security imaging object contour extraction method as described in any one of claims 1 to 7.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the sparse sampling three-dimensional security imaging object contour extraction method as described in any one of claims 1 to 7.

Images