A millimeter wave SAR imaging enhancement method

Abstract

The application provides a millimeter wave SAR imaging enhancement method, and belongs to the technical field of millimeter wave synthetic aperture radar imaging and industrial nondestructive testing, which can at least partially solve the problem that in the prior art, real defects and sidelobes, ghosting, virtual focus and other pseudo-aliasing cannot be distinguished due to the superposition of multiple structure responses in large-scale millimeter wave SAR imaging. The application comprises: obtaining millimeter wave echo data and generating a complex SAR image retaining phase information; generating a basic shape physical response library containing a main lobe and a pseudo-image label according to the structure characteristics of the object to be detected and the imaging geometry; initializing a candidate component set in the complex SAR image; obtaining soft contribution weights by using an alternating iterative solution based on a complex domain additive response model; determining real support confidence by comprehensively considering the soft contribution weights, physical support distance, phase consistency and pseudo-image label, retaining real structures and suppressing pseudo-image pixels. The application realizes interpretable enhancement of large-scale millimeter wave SAR images.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave synthetic aperture radar imaging and industrial non-destructive testing technology, specifically to a millimeter-wave SAR imaging enhancement method. Background Technology

[0002] Millimeter waves possess characteristics such as short wavelength, high imaging resolution, penetration capability through some non-metallic materials, minimal susceptibility to lighting conditions, and no ionizing harm to the human body. In recent years, they have been widely applied in fields such as security inspection, non-destructive testing, industrial inspection, and near-field high-resolution imaging. Synthetic Aperture Radar (SAR) technology, by creating an equivalent large aperture through moving antennas or arrays, can further improve lateral resolution, making it suitable for high-precision inspection of boiler water-cooled walls, pipes, plates, welds, composite material structures, and the surfaces of large industrial equipment.

[0003] However, existing millimeter-wave SAR imaging technology is more geared towards small objects, local targets, or relatively isolated scattering points. In large-scale inspection scenarios, the objects under test typically contain numerous edges, welds, curved pipe walls, holes, cracks, dents, and highly reflective components. The scattering responses of different structures are superimposed on each other in the imaging results. Due to the coherence of SAR imaging, the target echo not only appears as a real image on the main lobe but also includes side lobes, ghosting, defocus, and spatial diffusion responses caused by scanning trajectory, limited aperture, bandwidth limitations, surface curvature, occlusion, and multipath propagation. For large-scale complex targets, the true response of a local defect may be superimposed with artifacts generated by strong scattering structures in neighboring areas or at a distance, blurring the defect edges; some side lobes or ghosting may appear as bright lines resembling cracks, causing false detections. Simply improving the resolution of the SAR focusing algorithm cannot completely solve the structural aliasing problem in large-scale millimeter-wave imaging.

[0004] Existing techniques typically employ methods such as window weighting, sidelobe suppression, deconvolution, sparse reconstruction, low-rank decomposition, thresholding, edge detection, image filtering, deep learning denoising, or post-processing enhancement. While these methods are effective in noise suppression or local texture enhancement, they still have the following shortcomings: First, many methods process the final displayed intensity map, ignoring the complex phase information in millimeter-wave SAR images, making it difficult to handle aliasing caused by constructive and destructive interference. Second, traditional image segmentation or clustering methods often assume that a pixel mainly belongs to one region or category, while the intensity of a pixel in a real SAR image may be contributed by the main lobe, side lobes, ghosting, and defocus responses of multiple basic structures simultaneously. Third, conventional template matching or convolutional post-processing usually only focuses on the target's appearance and does not explicitly describe the complete SAR response generated by the basic shape under specific size, orientation, material, and imaging geometry, making it impossible to reliably determine whether a bright spot is a true support of the current structure or an artifact generated by other structures. Fourth, although deep learning methods can learn augmentation mappings from data, high-quality labeled data is limited in industrial non-destructive testing scenarios, and black-box models are difficult to explain the physical origin of defect candidate regions.

[0005] Therefore, how to fully utilize complex coherent information, explicitly model the physical response of the basic scattering structure, allow pixels to be contributed by multiple structures, and output interpretable artifact attribution results for large-scale millimeter-wave SAR imaging scenarios is a core problem that urgently needs to be solved in the field of large-scale millimeter-wave SAR imaging enhancement. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art and provide a millimeter-wave SAR imaging enhancement method.

[0007] To achieve the above objectives, the present invention provides a millimeter-wave SAR imaging enhancement method, comprising: Step S1: Acquire millimeter-wave echo data, perform amplitude and phase calibration and focusing processing on the millimeter-wave echo data, generate a complex SAR image that retains the amplitude and phase of each pixel, and record the imaging geometry; Step S2: Based on the structural characteristics of the object to be detected and the imaging geometry, a basic shape physical response library is generated. The response template of each basic shape in the basic shape physical response library is a complex response generated by the millimeter wave propagation model and the SAR focusing operator. The response template includes a main lobe region marker and an artifact region marker. Step S3: Detect candidate scattering features in the complex SAR image, match the candidate scattering features with the basic shape physical response library to generate a candidate component set, wherein each candidate component in the candidate component set has an initial shape parameter and an initial amplitude and phase parameter; Step S4: Based on the complex domain additive response model, the candidate component set is solved by alternating iteration. In the contribution estimation step, the soft contribution weight of the pixel from different components is estimated according to the predicted complex contribution of each component at each pixel. In the parameter update step, the shape parameters and amplitude and phase parameters of each component are updated so that the complex responses of all components are superimposed to approximate the complex SAR image. The contribution estimation step and the parameter update step are executed alternately until convergence. Step S5: Determine the true support confidence of each pixel based on the soft contribution weight, the distance between the pixel and the physical support region of the corresponding component, the phase consistency between the component's predicted response and the observed pixel, and the artifact region marker in the response template. Based on the true support confidence, retain the true structure pixels and suppress the artifact pixels, and output the enhanced SAR image.

[0008] Furthermore, in step S2, the basic shapes include point-like strong scatterers, linear cracks, edges, holes, welds, and arc surfaces, each of which is defined by shape parameters. Description, the shape parameters The response template includes center position, size, orientation angle, and equivalent reflection coefficient; the generation method of the response template includes one or a combination of analytical approximation, geometrical optics approximation, point scattering discretization, electromagnetic simulation, or experimental calibration; the artifact region marker distinguishes each pixel in the response template into main lobe region, side lobe region, ghost region, or out-of-focus region.

[0009] Furthermore, in step S3, scattering peaks, thin bright lines, and regions with local curvature changes are detected in the amplitude map of the complex SAR image as candidate scattering features, and background noise is eliminated by utilizing complex phase continuity; the candidate component set allows for overcomplete initialization, and the number of candidate components in the candidate component set is greater than the actual number of structures in the object to be inspected.

[0010] Further, in step S4, the complex domain additive response model is expressed as: ; in, Let be the complex SAR image vector. Number of pixels The total number of components, For the first Complex amplitude parameters of each component, For the first Each component in the imaging geometry Complex SAR response under the following conditions For the first The shape parameters of each component This includes noise and unmodeled components.

[0011] Furthermore, in the contribution estimation step, the first... The component in the first The prediction complex contribution at each pixel is The soft contribution weight The calculation formula is: ; in, The complex SAR response In the The component at each pixel For noise energy estimation, For stable terms; when the responses of multiple components overlap on the same pixel, multiple Both are greater than zero.

[0012] Furthermore, in the parameter update step, the shape parameter and amplitude / phase parameter of each component are updated with the following objective function: ; in, For regularization terms, The weight coefficients of the regularization term; the regularization term The shape parameters are constrained to fall within a physically feasible range.

[0013] Furthermore, step S4 also includes component management: if the total soft contribution of a component in the entire pixel range is lower than a preset contribution threshold or the magnitude of its amplitude and phase parameters is lower than the noise level, then the component is deleted; if the similarity of the shape parameters and responses of two components is higher than a preset similarity threshold, then the two components are merged and the parameters are re-estimated; if the residual region corresponding to a component presents multiple spatially separated peaks, then the component is split into multiple sub-components.

[0014] Furthermore, in step S5, the enhanced complex pixels and artifact residuals The calculation formula is: ; in, The complex SAR image vector The pixel value, For pixels Belongs to components The confidence level of the real support based on real physical support. For the first step S4 The component in the first The predicted complex contribution at each pixel; Preserve the true structural contribution. The artifacts that are identified.

[0015] Furthermore, step S1 also includes dividing the imaging area into multiple patches with overlapping boundaries, the overlap width being greater than the main lobe width and the side lobe width; after step S5, step S6 is also included: performing cross-pattern consistency determination on the component parameters in the overlapping area of ​​adjacent patches, when two components in adjacent patches are close in position, have the same direction and are connected at their length endpoints in global coordinates, the two components are merged into the same global structure, and weighted fusion is performed on the enhanced SAR image and artifact residual in the overlapping area.

[0016] Furthermore, step S5 also outputs a contribution map and an artifact map. The contribution map records the distribution of the soft contribution weight of each pixel on each component, and the artifact map records the location and source component of the suppressed artifact pixels. Step S5 also outputs explanation information for each defect candidate region. The explanation information includes the main contributing component of the candidate region, the true support confidence, the artifact confidence, and the phase consistency score.

[0017] The beneficial effects of this invention are as follows: This invention decomposes the real structural response and pseudo-effect response mixed together in a large-scale SAR image by using a complex domain additive response model, thereby reducing the occlusion and interference of strong scattering structures on nearby weak defects.

[0018] The soft contribution weighting mechanism of this invention allows a pixel to be contributed by multiple underlying structures simultaneously, which is suitable for explaining the response overlap phenomenon commonly seen in large-scale millimeter-wave SAR images and avoids the incorrect assignment caused by traditional hard classification methods.

[0019] This invention utilizes the main lobe region and artifact region markers in the basic shape physical response library to identify side lobes, ghosting, and out-of-focus areas. This not only enhances the image but also explains the source of the suppressed region, improving the interpretability of the detection results.

[0020] This invention maintains global structural continuity by determining cross-block consistency when processing large-scale blocks, which can reduce problems such as crack fracture, repeated alarms and boundary artifacts, making it suitable for large-area inspection tasks.

[0021] This invention outputs enhanced images, contribution maps, artifact maps, and defect candidate explanation information, which facilitates review by inspection personnel and the generation of inspection reports. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the millimeter-wave SAR imaging enhancement method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the basic shape physical response library generation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the alternating iterative solution process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the pixel-level soft contribution attribution and artifact removal process in an embodiment of the present invention; Figure 5 This is a schematic diagram of cross-tile fusion according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0024] The overall technical solution of the millimeter-wave SAR imaging enhancement method of the present invention will be described below.

[0025] See Figure 1 The millimeter-wave SAR imaging enhancement method of the present invention includes steps S1 to S5, and each step will be described in detail below.

[0026] Step S1 involves acquiring millimeter-wave echo data. Amplitude and phase calibration and focusing processing are performed on the millimeter-wave echo data to generate a complex SAR image that retains the amplitude and phase of each pixel, and the imaging geometry is recorded. The complex SAR image is the data foundation for all subsequent processing in this invention. Unlike traditional processes that only output intensity maps, this invention retains the complex values ​​of each pixel because the responses of different basic structures are approximately linearly superimposed in the complex domain, while nonlinear amplitude squared intersection terms appear in the intensity map. Direct decomposition on the intensity map can easily lead to incorrect attribution. The imaging geometry... Information including aperture position, frequency sampling, coordinate system transformation, and target area range is used as input conditions for generating the response library in subsequent step S2.

[0027] In this invention, a "complex SAR image" refers to image data containing both amplitude and phase information, obtained after range compression, phase compensation, and focusing processing of millimeter-wave echo data. The focusing processing can employ, but is not limited to, the following methods: back projection algorithm, frequency domain imaging algorithm, range migration algorithm, or near-field imaging algorithm. The commonality of these focusing methods lies in that they all improve lateral resolution by compensating for the propagation path difference between the antenna and the target, and all output focusing results that retain complex information. In various embodiments of this invention, different focusing algorithms can be selected based on computational resources and real-time requirements; the selection does not affect the execution logic of subsequent steps S2 to S5.

[0028] In a further embodiment, step S1 may further divide the imaging region into multiple tiles with overlapping boundaries. Tile division allows large imaging regions to be processed block by block with limited computing resources. The overlap width is preferably greater than the width of the main lobe and the width of the main side lobes, ensuring that cracks or edges near the boundaries appear completely in at least one tile. Each tile stores the transformation relationship between its local and global coordinate systems, facilitating subsequent cross-tile component merging and global stitching. The tile management module also records the imaging geometry, local noise estimation, background level, and boundary overlap relationship corresponding to each tile.

[0029] Step S2 involves generating a basic shape physical response library based on the structural characteristics of the object to be detected and the imaging geometry. The response template for each basic shape in the basic shape physical response library is a complex response generated by the millimeter-wave propagation model and the SAR focusing operator. The response template includes main lobe region markers and artifact region markers. The basic shape physical response library is the core component of this invention, transforming the imaging enhancement problem into an additive inverse problem under physical constraints.

[0030] In this invention, the "basic shape physical response library" refers to the collection of templates that generate corresponding complex SAR response templates for the basic scattering structures that may appear in the object to be detected, under a given imaging geometry, through a millimeter-wave propagation physical model and SAR focusing operators, and are then compiled into a queryable set of templates. "Physical response" means that this template is not a regular image appearance template, but a complex response determined by the geometric parameters of the basic shape and the imaging geometry, obtained after millimeter-wave propagation and SAR focusing processing. A response template simultaneously includes the main lobe response near the target's actual support location, as well as side lobes, ghosting, and out-of-focus responses caused by finite aperture, window function, scan trajectory, and geometric occlusion. "Main lobe region markers" indicate the response region near the target's actual physical support, and "artifact region markers" indicate the response regions corresponding to side lobes, ghosting, and out-of-focus. In subsequent step S5, when determining whether a bright spot is a real structure, the system will combine these markers, rather than relying solely on the brightness threshold.

[0031] Furthermore, the generation of the basic shape physical response library can be achieved, but is not limited to, the following methods: analytical approximation, geometrical optics approximation, point scattering discretization, full-wave electromagnetic simulation, or experimental calibration. The commonality of these generation methods lies in the fact that they all use basic shape parameters and imaging geometry as input, and all obtain a complex response template containing the main lobe and artifact regions through forward propagation calculations. Analytical approximation is suitable for basic shapes whose scattering behavior can be described by a closed-form formula, and has the fastest computational speed; geometrical optics approximation is suitable for cases where the target size is much larger than the wavelength, calculating the scattering field through physical optics surface element integration; point scattering discretization discretizes the basic shape into multiple equivalent scattering points and accumulates their echoes point by point, suitable for arbitrary shapes; full-wave electromagnetic simulation obtains the accurate scattering response by numerically solving Maxwell's equations, offering the highest accuracy but also the highest computational cost; experimental calibration obtains the template through actual measurements of known standard parts, requiring no modeling but limited by the type of standard parts. In various embodiments of this invention, one or a combination of the above methods can be selected and used according to the complexity of the object and the accuracy requirements.

[0032] Step S3 involves detecting candidate scattering features in the complex SAR image, matching these features with the basic shape physical response library, and generating a candidate component set. This candidate component set provides an initial starting point for subsequent iterative solving. This step extracts potential scattering structures from the complex SAR image and assigns initial shape parameters and initial amplitude and phase parameters to each candidate structure. Since there are many artifacts in large-scale complex targets, the initialization phase allows for an over-complete set of candidate components, meaning the number of candidate components exceeds the actual number of structures. This is gradually adjusted through component management in subsequent iterative solving. The detection methods for the candidate scattering features can include, but are not limited to, the following: scattering peak detection based on amplitude thresholds, thin bright line detection based on aspect ratio, edge detection based on gradients, or shape detection based on curvature.

[0033] Step S4 involves applying an alternating iterative solution to the candidate component set based on the complex domain additive response model. This alternating iterative solution is the core computational process of this invention. In this invention, the "alternating iterative solution" borrows the alternating idea of ​​"first estimating the latent contribution, then updating the model parameters" from the Expectation-Maximization (EM) algorithm. However, this invention estimates the superimposed soft contribution rather than mutually exclusive class assignments. Therefore, this invention refers to it as the EM-like (class-based Expectation-Maximization) method.

[0034] Furthermore, the alternating iterative solution is a complex-domain additive inversion solution method that iterates between the contribution estimation step and the parameter update step until convergence. The input to this method is a complex SAR image, a set of candidate components, imaging geometry, and a physical response library of basic shapes. In the contribution estimation step, the system calculates the non-mutually exclusive energy attribution ratio of each component at each pixel, i.e., the soft contribution weight. In the parameter update step, the system updates the parameters of each component with the goal of minimizing the reconstruction error of the complex-domain additive model. The output of the alternating iteration is the converged soft contribution weight and the optimized parameters of each component. Parameter updates can be performed using, but are not limited to, the following methods: gradient descent, local grid search, matching pursuit, or nonlinear least squares (Levenberg-Marquardt algorithm). The commonality of these methods is that they all alternate between the contribution estimation step and the parameter update step, and all use the reconstruction error of the complex-domain additive response model as the optimization objective.

[0035] In the alternating iterative solution, the contribution estimation step estimates the soft contribution weights of pixels from different components based on the predicted complex contributions of each component at each pixel. These soft contribution weights are not mutually exclusive class probabilities in traditional clustering, but rather pixel energy attribution proportions; when the main lobes or artifacts of multiple components overlap in the same pixel, the soft contribution weights corresponding to multiple components can be simultaneously large. This step allows the same pixel to be interpreted by multiple components simultaneously, especially suitable for cases where strong scattering edge sidelobes overlap with crack main lobes. In the parameter update step, the shape and amplitude / phase parameters of each component are updated so that the complex responses of all components are superimposed to approximate the complex SAR image. The contribution estimation step and the parameter update step are executed alternately until convergence.

[0036] The EM-like method of this invention differs from ordinary EM clustering. Ordinary EM clustering typically assumes that each sample comes from a component of a probability distribution, while in this invention, each component is a physical response function, and a pixel may contain the superimposed contributions of multiple components. Therefore, the contribution estimation step of this invention does not calculate mutually exclusive class probabilities, but rather estimates pixel-level energy attribution; the parameter update step does not update the mean and variance of the probability distribution, but rather updates the physical parameters of the basic shape and the SAR complex response. This difference makes this method more suitable for the coherent superposition nature of millimeter-wave SAR imaging.

[0037] Step S5 determines the true support confidence of each pixel based on the soft contribution weight, the distance between the pixel and the corresponding component's physical support region, the phase consistency between the component's predicted response and the observed pixel, and the artifact region marker in the response template. Based on the true support confidence, pixels representing the true structure are retained while artifact pixels are suppressed, and the enhanced SAR image is output. This step integrates four types of information to determine the affiliation of each pixel: First, the soft contribution weight of the pixel to the component; a larger weight indicates a greater contribution from the component's response. Second, the distance between the pixel and the component's true support region in physical coordinates; a smaller distance indicates the pixel is closer to the true structure location. Third, the phase consistency between the component's predicted response and the observed pixel; a higher phase consistency indicates a better match between the predicted and observed phases. Fourth, the marker in the response template indicating whether the pixel corresponds to a main lobe region or an artifact region. Component responses with high true support confidence are retained by the system; responses identified as sidelobes, ghosting, or out-of-focus are suppressed in the enhanced image, while their location and origin are preserved in the artifact image.

[0038] The above is a description of the overall technical solution of the present invention. The following detailed explanation of each step is provided through specific embodiments.

[0039] Example 1 This embodiment uses boiler water-cooled wall crack detection as an application scenario. The water-cooled wall consists of parallel steel tubes with an outer diameter of... Wall thickness The tube spacing is 76mm, and the tube material is 12Cr1MoVG. The millimeter-wave scanning device is mounted on a moving platform and moves along the surface of the water-cooled wall according to a preset trajectory. The scanning frequency range is 75GHz to 110GHz (W band), with a frequency step of 50MHz, a total of 701 frequency points, and a scanning aperture length of 500mm.

[0040] Step S1: Acquire millimeter-wave echo data, perform amplitude and phase calibration and focusing processing on the millimeter-wave echo data, generate a complex SAR image that retains the amplitude and phase of each pixel, and record the imaging geometry.

[0041] The millimeter-wave scanning device employs a single-transmitter, single-receiver equivalent synthetic aperture array structure. The antenna moves along the surface of the water-cooled wall in 5mm steps to form the equivalent synthetic aperture. The system records the antenna coordinates at each sampling position. Antenna normal attitude, sweep start and end frequencies to Frequency step Transmission power and receive gain The system simultaneously records the position information of the moving platform and the normal geometric information of the water-cooled wall surface provided by the encoder.

[0042] After acquisition, amplitude calibration, phase calibration, background subtraction, and system delay compensation are performed on the raw echo. Background subtraction is achieved through empty-field measurement: the background echo is acquired when there is no object under test. The actual echo Subtracting the background echo yields the target echo. This process removes system clutter and environmental reflections. Amplitude and phase calibration is performed using a standard trihedral corner reflector: the corner reflector is placed at a known location, and the ratio of its theoretical scattering cross-section to the measured echo is used to correct the system's frequency response amplitude deviation; the difference between its theoretical echo phase and the measured echo phase is used to correct the system's phase shift. If there are changes in platform speed or trajectory errors during the scanning process, the system repositions the sampling point coordinates based on encoder data.

[0043] The calibrated echo data is then focused using a back projection algorithm. The back projection algorithm performs focusing on each pixel position. , cumulative all Each sampling location and The contribution of the echo signal at each frequency point after phase compensation: ; in For the first Each sampling location to pixel distance, At the speed of light, For the first The frequency of each frequency point. After focusing, a complex SAR image is generated, retaining the amplitude and phase of each pixel. The complex SAR image vector is denoted as... , The number of pixels is used. In this embodiment, the imaging resolution is approximately 3mm (distance direction) × 5mm (azimuth direction).

[0044] Simultaneously record imaging geometry The imaging geometry Including aperture position sequence Frequency sampling parameters Coordinate system transformation matrix and target area range.

[0045] In this embodiment, the imaging area is the 2m × 1m surface of the water-cooled wall. The imaging area is divided into 16 overlapping patches, each 600mm × 350mm in size, with an overlap width of 100mm between adjacent patches, greater than the main lobe width (approximately 3mm) and the main side lobe width (approximately 30mm). Each patch stores the position and orientation transformation relationship of its local coordinate system origin in the global coordinate system. The patch management module records the imaging geometry, local noise variance, background scattering level, and boundary overlap relationship with adjacent patches for each patch. For periodic tubular structures like water-cooled walls, the system uses the pipe spacing of 76mm, pipe diameter of 60mm, and weld orientation as priors to associate adjacent patches on the same pipe wall or the same weld.

[0046] Step S2: Generate a basic shape physical response library based on the structural characteristics of the object to be detected and the imaging geometry.

[0047] For water-cooled wall crack detection scenarios, the structural characteristics of the object to be inspected include the pipe wall arc surface, weld edges, axial cracks, circumferential cracks, and localized strong scattering points. The basic shapes include the following types: Point-like strong scattering bodies: describe solder joints, corners, or localized metal protrusions. Shape parameters Including the central location and equivalent reflection coefficient .

[0048] Short linear cracks: Describe microcracks with a length of less than 20 mm. Shape parameters Including the central location ,length ,width Direction angle Equivalent reflection coefficient and surface height difference .

[0049] Long linear cracks: Describe through cracks with a length greater than or equal to 20 mm. Shape parameters Similar to short linear cracks, the length range extends to 200 mm.

[0050] Tube wall arc surface segment: Describes the cylindrical arc surface of the water-cooled wall tube. Shape parameters Including axis position ,radius Arc length Axial range and surface reflectance .

[0051] Weld edge: Describes the weld seam between pipe walls and its edge transition zone. Shape parameters Including the central location ,length ,width Direction angle and weld height .

[0052] Holes: Describes circular or elliptical defects in the pipe wall. Shape parameters Including the central location Long axis short axis Direction angle and depth .

[0053] For each basic shape and each set of shape parameters, the response library generation module generates a response based on the imaging geometry. The SAR response template of this basic structure is calculated. In this embodiment, the response template is generated using a point scattering discretization method. Taking a linear crack as an example: the crack is uniformly discretized along its length as follows: One equivalent scattering point, ,in The wavelength corresponding to the center frequency (approximately 3.2 mm); based on the position of each scattering point. and equivalent scattering cross section Calculate its echo signal under the current antenna array and frequency sweep parameters. The echoes from all scattered points are summed, and then back-projection focusing is performed to obtain the complex response template of the linear crack. .

[0054] For the same basic shape, template families are generated under different sizes, orientations, and positions. The response template includes the main lobe response near the actual support location of the target, as well as the side lobe response caused by finite aperture and window function, the ghosting response caused by multipath propagation, and the out-of-focus response caused by surface curvature. Each pixel in the response template is classified as a main lobe region, side lobe region, ghosting region, or out-of-focus region. The main lobe region is determined as follows: within a range extending one resolution cell in both the range and azimuth directions, centered on the geometric support area of ​​the target, the response energy exceeds the peak energy. The region marked is the main lobe region. The side lobe regions are the areas on either side of the main lobe whose attenuation distribution follows the imaging system's point spread function; their boundaries decrease from the response energy to the peak energy. The location is determined. The ghosting region is a bright spot area far from the main lobe caused by multipath or mirror effects. The out-of-focus region is a blurry response region caused by focus shift due to surface curvature.

[0055] In this embodiment, the response library is pre-generated offline. There are 50 templates for point-like strong scatterers (covering 10 reflection coefficient levels, with 5 position offsets per level), 2000 templates for linear cracks (covering lengths from 5mm to 200mm in 20 levels, azimuth angles from 0° to 180° in 18 levels, and width levels), 200 templates for pipe wall arc segments (covering different arc lengths and axial ranges), 500 templates for weld edges (covering different lengths and weld heights), and 100 templates for holes. The total number of templates is approximately 2850.

[0056] Step S3: Detect candidate scattering features in the complex SAR image, match the candidate scattering features with the basic shape physical response library, and generate a candidate component set.

[0057] Amplitude plot of the complex SAR image In this process, the system performs the following three types of feature detection. First, scattering peak detection: calculating the local background mean for each pixel. and standard deviation (centered on this pixel) (Statistics within a pixel window), excluding amplitude values ​​exceeding The first step is to identify scattering peaks in the pixels. The second step is to detect elongated bright lines: perform morphological skeleton extraction on the amplitude image, and mark skeleton segments with an aspect ratio greater than 5:1 and more than 10 consecutive pixels as candidates for elongated bright lines. The third step is to detect regions with local curvature changes: calculate the second spatial derivative of the amplitude image, and mark regions with curvature changes exceeding a preset curvature threshold as candidates.

[0058] Simultaneously, the phase continuity of complex SAR images is utilized to eliminate isolated random noise bright spots: for each candidate scattering peak, the average phase difference between it and its 8 neighboring pixels is calculated. ,like If so, the bright spot is determined to be noise and eliminated.

[0059] Each detected candidate scattering feature is coarsely matched with the basic shape physical response library. The coarse matching employs a normalized complex cross-correlation method: a local complex image patch is extracted centered on the candidate scattering feature region; the normalized complex cross-correlation value between this image patch and each template in the response library is calculated; the template type with the highest cross-correlation value is selected as the initial shape type for that candidate, and the parameters of this template are used as the initial shape parameters. Initial amplitude and phase parameters It is determined by the complex value of the peak pixel in the candidate region.

[0060] In the boiler water-cooled wall scenario, the initialization module also utilizes structural priors. Pipe wall edges and welds are typically distributed axially, while cracks may be distributed axially, circumferentially, or near the weld. The system assigns higher initial weights to slender candidates that conform to the prior directions and lower initial weights to isolated bright spots far from physical supports. These weights are not used as the final judgment criterion but only to improve the iteration convergence speed.

[0061] In this embodiment, for the 2m×1m water-cooled wall imaging region, step S3 detects approximately 120 candidate scattering features, generating the candidate component set containing 120 candidate components. The number of candidate components in the candidate component set is greater than the actual number of structures on the water-cooled wall surface (approximately 50 real structures), thus allowing for overcomplete initialization.

[0062] Step S4: Solve the candidate component set using an alternating iterative approach based on the complex field additive response model.

[0063] The complex domain additive response model is expressed as: ; in, Let be the complex SAR image vector. This refers to the number of pixels in a single tile in this embodiment (approximately 600,000). This represents the total number of candidate components (initially 120). For the first Complex amplitude parameters of each component, For the first Each component, based on the corresponding template in the basic shape physical response library, is in the imaging geometry The generated complex SAR response vector, For the first The shape parameters of each component This includes noise and unmodeled components.

[0064] The objective of the alternating iterative solution is to minimize the additive inverse problem with regularization terms: ; in, Used to penalize excessive components and prevent noise from being interpreted as real structure; This is a regularization term used to constrain shape parameters to fall within the physically feasible range, including size range constraints (crack width not exceeding 5mm), orientation range constraints (the radius of the tube wall arc surface should be consistent with the diameter of the water-cooled wall tube), and spatial range constraints (the component position should not jump out of the current block's valid area). This represents the weighting coefficient of the regularization term. In this embodiment... , .

[0065] Alternating iterative solution involves contribution estimation and parameter update steps, which are executed alternately.

[0066] Contribution estimation step: Generate the complex response of each component based on the current component parameters and superimpose them to obtain the current predicted image. Calculate the first... The component in the first Predicted complex contribution at each pixel ,in For response vector The Each component. The soft contribution weight. The calculation formula is: ; in, For noise energy estimation, the variance of regions with no obvious structure in the image (the defect-free area in the center of the pipe wall) is estimated in this embodiment. . To stabilize the term and prevent the denominator from being zero, this embodiment... When the responses of multiple components overlap on the same pixel, multiple It is also greater than zero. This weight is not the probability of mutually exclusive categories in traditional clustering, but rather the pixel energy attribution ratio.

[0067] Parameter update steps: Fix or utilize the soft contribution weights to update the shape and amplitude parameters of each component, so that the complex responses of all components are superimposed to approximate the complex SAR image. ; For complex amplitude When the shape parameters are fixed, the solution is obtained by complex least squares: the response vectors of all components are combined into a matrix. ,but ,in This is the conjugate transpose. This is the identity matrix. For nonlinear parameters such as position, size, orientation, and attitude, this embodiment uses gradient descent to update them: the gradient of the objective function with respect to each shape parameter is calculated, using the learning rate... Update parameters along the negative gradient direction. During the update process, the constraint component parameters fall within the physically feasible range: crack width. Pipe wall arc radius (Same as water-cooled wall tube diameter), crack length The component position does not jump out of the current valid area of ​​the tile.

[0068] Step S4 also includes component management. Component management is performed once every 5 iterations and includes three operations. Pruning operation: If the total soft contribution of a component across the entire pixel range... Below the preset contribution threshold or its complex amplitude Below the noise standard deviation If so, delete the component and reduce the total number of components. Decrease by one. Merge operation: If two components and Same shape type, distance between positions Angular difference And the normalized cross-correlation of the response Then, the two components are merged into one component, and the shape and amplitude parameters of the merged component are re-estimated using a weighted average. Splitting operation: Calculate the residual plot for each component. ,right Performing connected component analysis, if the residual region exhibits two or more spatially separated peaks (the number of connected components is greater than 1 and each peak exceeds [a certain threshold]), then [the analysis is successful]. If the component is split into multiple sub-components, each sub-component corresponds to a residual peak, and the initial position of the sub-component is set to the center of the corresponding peak.

[0069] In this embodiment, the iteration termination condition is the relative change in the objective function between adjacent iterations. The iteration count reached 200 times. After about 80 iterations, the initial 120 candidate components converged to 52 components after pruning, merging and splitting. Among them, there were 26 pipe wall arc surface segments, 8 weld edges, 12 linear cracks, 4 point strong scatterers and 2 holes.

[0070] Step S5: Determine the true support confidence of each pixel based on the soft contribution weight, the distance between the pixel and the physical support region of the corresponding component, the phase consistency between the component's predicted response and the observed pixel, and the artifact region marker in the response template. Based on the true support confidence, retain the true structure pixels and suppress the artifact pixels, and output the enhanced SAR image.

[0071] For each pixel and each component The system comprehensively considers the following four types of information to determine the confidence level of true support. .

[0072] First, pixels For components The soft contribution weight .

[0073] Second, pixels In physical coordinates and components Real support area distance Actual support area For components The physical space occupied by geometric parameters. Taking a linear crack as an example, its actual support region is centered on the crack center and has a length of... Width is A rectangular area.

[0074] Third, components In pixels Predictive Complexity Contribution With observed pixels Phase consistency between .

[0075] Fourth, pixels In components The corresponding region type tag in the response template . Indicates the main lobe region. Indicates the artifact region (side lobe, ghosting, or out-of-focus).

[0076] True support confidence level The above four types of information are used to determine the following: ; in, For distance attenuation parameters, in this embodiment When pixels Located in the component The main lobe region ( ), close to the physical support area ( Small), high soft contribution weight ( When the phase is large and the phase consistency is good, A value close to 1 indicates that the pixel is a component. The true structure of the pixel. When the pixel is located in the artifact region ( )hour, This pixel was identified as an artifact pixel.

[0077] Enhanced complex pixels and artifact residuals The calculation formula is: ; Preserve the true structural contribution. For residual terms identified as sidelobes, ghosting, out-of-focus, or unmodeled interference.

[0078] This embodiment also outputs a contribution map and an artifact map. The contribution map records the soft contribution weight of each pixel. The distribution of each component is presented in a color-coded manner, allowing inspectors to see which components primarily contribute to a particular pixel. The artifact map records the location and source component number of suppressed artifact pixels, using different gray levels to indicate three artifact types: sidelobes, ghosting, and out-of-focus. The explanatory information for each defect candidate region includes: the global coordinates, length or area, orientation, tile number, corresponding basic shape type, main contributing component number, mean true support confidence score, mean artifact confidence score, phase consistency score, and suggested review level.

[0079] Step S6: Perform cross-tile consistency determination and weighted fusion on adjacent tiles.

[0080] For component parameters within overlapping areas of adjacent tiles, the system performs cross-tile consistency determination. When two crack components in adjacent tiles are less than 10mm apart in global coordinates, have an angular difference of less than 10°, and a length endpoint distance of less than 15mm, the two components are merged into the same global crack and assigned a unified component number in global coordinates. If a bright line in a tile lacks true support continuity in adjacent tiles but is consistent with the sidelobe template of a distant strong scatterer, it is marked as a local artifact rather than a new defect.

[0081] Weighted fusion is performed on the enhanced SAR image and artifact residuals within the overlapping region. Fusion weights. From pixels Distance from the center of the tile Sure: ,in The attenuation parameter is the center distance of the image patch. In this embodiment... For the same physical location, enhanced pixels from different patches are weighted and averaged according to their respective fusion weights to obtain a global enhanced SAR image. For repeatedly detected candidate defects, the system merges their coordinate range and confidence level to avoid the same crack triggering repeated alarms in multiple patches.

[0082] Finally, the system outputs a global enhanced SAR image, a true structure support map, a contribution map, an artifact map, and a list of defect candidate regions. For each defect candidate region, the output includes global coordinates, length or area, orientation, location in the map tile, corresponding basic shape type, main contributing components, true support confidence, artifact confidence, and suggested review level.

[0083] In the boiler water-cooled wall crack detection of this embodiment, the traditional intensity map processing method has a false detection rate of 15% due to misclassifying weld edge lobes as cracks, and a false detection rate of 12% due to small cracks being obscured by the false focus response. After adopting the method of this invention, the false detection rate is reduced to 3%, and the false detection rate is reduced to 4%. An axial crack (25mm long, 0.3mm wide) located near the weld edge is completely obscured by weld edge lobes in the traditional intensity map. This invention attributes the weld edge lobes to weld edge components and marks them as artifacts through complex domain decomposition, while preserving the true support response of the crack components, making the crack clearly visible in the enhanced image. The processing time for a single block is approximately 12 seconds, and the total processing time for all 16 blocks is approximately 4 minutes.

[0084] Example 2 This embodiment uses pipeline weld inspection as an application scenario. The pipeline under inspection is the main steam pipeline of a power plant, and the pipeline outer diameter is... Wall thickness The pipe material is P91. The weld is a circumferential butt weld with a weld reinforcement of approximately 2mm to 3mm. A millimeter-wave scanning device is mounted on a ring-shaped guide rail on the outer surface of the pipe, scanning circumferentially. The scanning frequency range is 75GHz to 110GHz (W band), with a frequency step of 50MHz, for a total of 701 frequency points. The scanning angle range of the ring-shaped guide rail is 0° to 360°, with an angle step of 0.5°, for a total of 720 sampling positions. The scanning aperture is half the circumference of the pipe, approximately 344mm.

[0085] Step S1: Acquire millimeter-wave echo data, perform amplitude and phase calibration and focusing processing on the millimeter-wave echo data, generate a complex SAR image that retains the amplitude and phase of each pixel, and record the imaging geometry.

[0086] The millimeter-wave scanning device employs a single-transmitter, single-receiver structure. The antenna is mounted on a slider on a circular guide rail and moves circumferentially along the outer surface of the pipe in 0.5° increments. The system records the antenna's angular position in the pipe coordinate system at each sampling location. ( Radial distance from the antenna to the pipe surface The system records the start and end frequencies of the frequency sweep and the transmit power. Since the pipe surface is a cylindrical curved surface, the system also records the pipe's outer diameter. The position of the pipeline axis is used for surface phase compensation in subsequent focusing processing.

[0087] After acquisition, amplitude and phase calibration and background subtraction are performed on the original echo. The background subtraction method is the same as in Example 1, where the background echo is obtained through open-field measurement and subtracted from the target echo. Amplitude and phase calibration is performed by installing a standard metal block (a rectangular metal boss of 10mm × 10mm × 2mm) at a known location on the pipe surface, and the system frequency response is corrected by using the ratio of the theoretical scattering cross section of the standard metal block to the measured echo.

[0088] The calibrated echo data is focused using a back projection algorithm in cylindrical coordinates. Unlike the back projection in Cartesian coordinates in Example 1, the back projection algorithm in this example needs to correct the propagation path length from the antenna to each pixel position based on the pipe curvature. For the pipe surface after unfolding, the coordinates are... pixels ( For the circumferential arc length, (where axial position is used), its coordinates in three-dimensional space are: The propagation path length from the antenna to the pixel Calculated using three-dimensional Euclidean distance. After focusing, a complex SAR image is generated, preserving the amplitude and phase of each pixel, with a pixel resolution of approximately 3 mm in the range direction and approximately 2 mm in the circumferential direction.

[0089] Simultaneously record imaging geometry The imaging geometry Including the angle sequence of each sampling position on the circular guide rail Frequency sampling parameters, pipe outer diameter The relationship between the pipe axis position and the transformation from the cylindrical coordinate system to the unfolded coordinate system.

[0090] In this embodiment, the imaging area is the outer surface of the pipe near the weld. The annular region, when unfolded, is The rectangular region was divided into 8 tiles with overlapping boundaries, each tile being [size missing]. The overlap width between adjacent blocks is 60mm. Each block stores its position in the pipe unfolding coordinate system.

[0091] Step S2: Generate a basic shape physical response library based on the structural characteristics of the object to be detected and the imaging geometry.

[0092] For pipeline weld inspection scenarios, the structural characteristics of the object to be inspected include circumferential weld reinforcement, weld edge transition zone, internal weld cracks, heat-affected zone cracks, and the arc surface of the pipe wall. The basic shapes include the following types.

[0093] Weld reinforcement: Describes the portion of the weld surface that protrudes above the base material. (Shape parameter) Including the central location Circular span axial width Remaining height and surface reflectance .

[0094] Weld edge: Describes the transition area between the weld and the base metal. Shape parameters Including the central location Circular span , transition zone width Direction angle and reflection coefficient .

[0095] Cracks within the weld: Describes cracks distributed along the weld direction (circumferential). Shape parameters Including the central location ,length ,width Direction angle (Approximately 0° or 360° circumferential) and equivalent reflection coefficient .

[0096] Heat-affected zone cracks: Describe cracks distributed perpendicular to the weld direction (axial direction). Shape parameters Similar to cracks within the weld, the direction angle Approximately 90° axially.

[0097] Pipe wall arc surface segment: Describes the cylindrical arc surface of the pipe surface. Shape parameters Including axis position Pipe radius Arc length and axial range .

[0098] Point-like strong scattering bodies: describe porosity, slag inclusions, or localized protrusions on the weld surface. Shape parameters. Including the central location and equivalent reflection coefficient .

[0099] The response template is generated in the same way as in Example 1, using a point scattering discretization method. Due to the large curvature of the pipe surface (radius of curvature is...), In template generation, the influence of the pipe wall arc surface on the direction of scattered wave propagation needs to be considered: the normal direction of each discrete scattering point changes with the circumferential position, and its equivalent scattering cross section needs to be corrected according to the angle between the incident angle and the surface normal. For template generation of weld reinforcement, the reinforcement part needs to be modeled as a stepped protrusion, and the joint echo of the scattering points at the upper and lower edges of the step and the top surface needs to be calculated.

[0100] In this embodiment, the response library is pre-generated offline. It includes 200 weld reinforcement templates (covering different reinforcement heights and widths), 300 weld edge templates (covering different transition zone widths), 800 weld internal crack templates (covering lengths from 5mm to 100mm and widths from 0.1mm to 2mm), 600 heat-affected zone crack templates, 150 pipe wall arc surface segment templates, and 50 point-like strong scattering body templates. Each template in the response library contains main lobe region markers and artifact region markers. The total number of templates is approximately 2100.

[0101] Step S3: Detect candidate scattering features in the complex SAR image, match the candidate scattering features with the basic shape physical response library, and generate a candidate component set.

[0102] The detection method for candidate scattering features is the same as in Example 1. Scattering peaks, thin bright lines, and regions of local curvature variation are detected in the amplitude plot of the complex SAR image, and noise bright spots are eliminated using phase continuity. Since the background scattering (inherent echo of the weld reinforcement) in the pipe weld area is stronger than the background scattering from the water-cooled wall in Example 1, the threshold for scattering peak detection in this example is adjusted to the local background mean plus four times the standard deviation. The coarse matching method and the way initial parameters are determined are the same as in Example 1.

[0103] In the pipeline weld scenario, the system utilizes prior knowledge of the weld location. Since the location of the circumferential weld is known (located at the axial center of the imaging area), the system focuses on the weld centerline. Slender candidate templates within the range are preferentially matched to weld crack templates located at the weld centerline. to Candidate templates for heat-affected zone cracks within the specified range are given priority matching.

[0104] In this embodiment, step S3 detects approximately 80 candidate scattering features and generates the candidate component set containing 80 candidate components.

[0105] Step S4: Solve the candidate component set using an alternating iterative approach based on the complex field additive response model.

[0106] The additive response model in the complex domain is the same as in Example 1. In this example, This represents the number of pixels in a single tile (approximately 230,000). Initially 80. Noise energy estimation. Stable terms Component number penalty coefficient Regularization weights .

[0107] In the contribution estimation step, the soft contribution weight is calculated in the same way as in Example 1.

[0108] In the parameter update step, this embodiment uses the nonlinear least squares method (Levenberg-Marquardt algorithm) to update the shape parameters. Compared with the gradient descent method used in Embodiment 1, the Levenberg-Marquardt algorithm, by adaptively switching between the Gauss-Newton method and gradient descent, exhibits higher convergence speed and more stable convergence behavior in curved surface scenarios. Complex amplitude parameters The update method is the same as in Example 1, solved using complex least squares. During the update process, the constraint component parameters fall within the physically feasible range: crack width... Pipe arc radius Crack length .

[0109] The component management rules are the same as in Example 1, performing pruning, merging, and splitting operations once every 5 iterations. Pruning threshold. The threshold for merging location distance is 3mm, the threshold for merging direction angle difference is 5°, and the threshold for merging cross-correlation is 0.9.

[0110] In this embodiment, the iteration termination condition is the same as in Embodiment 1. After approximately 60 iterations, convergence occurs, with 80 initial candidate components converging to 38 components, including 4 weld reinforcement components, 6 weld edge components, 8 weld internal crack components, 5 heat-affected zone crack components, 12 pipe wall arc surface segments, and 3 point-like strong scattering bodies.

[0111] Step S5: Determine the true support confidence, retain true structure pixels and suppress artifact pixels, and output the enhanced SAR image.

[0112] The calculation method for the true support confidence is the same as in Example 1, with the distance attenuation parameter... (In the scenario of pipe welds, the defect scale is relatively small, so the distance attenuation parameter is adjusted accordingly to improve the sensitivity for detecting small defects.) Enhanced complex pixels and artifact residuals The calculation formula is the same as in Example 1.

[0113] This embodiment also outputs contribution maps, artifact maps, and explanation information for defect candidate regions. The generation methods for contribution maps and artifact maps are the same as in Embodiment 1. The explanation information for each defect candidate region includes: the position of the candidate region in the pipe unfolding coordinate system (circumferential arc length coordinates and axial coordinates), length or area, direction (circumferential or axial), corresponding basic shape type, main contributing component number, mean confidence score of true support, mean confidence score of artifacts, and phase consistency score.

[0114] Step S6: Perform cross-tile consistency determination and weighted fusion on adjacent tiles.

[0115] The method for cross-tile consistency determination and weighted fusion is the same as in Example 1. Since the pipe is a closed cylindrical surface, the last tile is adjacent to the first tile in the circumferential direction. The system performs a circumferential closed-loop matching to ensure the continuity of circumferential weld cracks at the 0° / 360° junction. The calculation method for the fusion weight is the same as in Example 1. .

[0116] In pipeline weld inspection, traditional strength map processing methods have a false detection rate of 18% for misclassifying weld reinforcement side lobes as cracks, and a false negative rate of 14% for microcracks within the weld (length less than 10 mm and width less than 0.2 mm). Using the method of this invention, the false detection rate is reduced to 4%, and the false negative rate is reduced to 5%. The strong side lobe response of the weld reinforcement is accurately attributed to the weld reinforcement component and marked as an artifact, ensuring that two axial microcracks (lengths of 6 mm and 8 mm, respectively) located in the heat-affected zone are preserved and clearly displayed in the enhanced image. The processing time for a single block is approximately 8 seconds, and the total processing time for all 8 blocks is approximately 2 minutes.

[0117] Example 3 This embodiment uses security inspection imaging as an application scenario. The object being inspected is a millimeter-wave full-body scan imaging of the human body surface and items carried inside clothing. The millimeter-wave scanning device uses a planar linear array, with the array direction vertical, an array length of 2m, containing 400 array elements, and an element spacing of 5mm. The array moves horizontally in a mechanical scanning manner, with a horizontal scanning range of 0.8m, a horizontal step of 5mm, and a total of 160 sampling positions. The scanning frequency range is 24GHz to 30GHz (Ka band), with a frequency step of 25MHz, and a total of 241 frequency points. The distance from the antenna to the object being inspected is approximately 0.5m to 1.0m.

[0118] Step S1: Acquire millimeter-wave echo data, perform amplitude and phase calibration and focusing processing on the millimeter-wave echo data, generate a complex SAR image that retains the amplitude and phase of each pixel, and record the imaging geometry.

[0119] The millimeter-wave scanning device simultaneously acquires echo data from 400 array elements at each horizontal scanning position. The system records the position of each array element in the global coordinate system (horizontal coordinates). and vertical coordinates , , The antenna array elements' positional accuracy is guaranteed by factory calibration, with a positional error of less than 0.1 mm.

[0120] After acquisition, amplitude and phase calibration and background subtraction are performed on the raw echoes. Amplitude and phase calibration is completed through the built-in reference channel: the system automatically switches to the internal reference signal before each scan, measures and compensates for the amplitude and phase differences between each array element. Background subtraction is achieved through unmanned field measurements.

[0121] The calibrated echo data is focused using a frequency domain imaging algorithm to meet the real-time requirements of security inspection scenarios. The frequency domain imaging algorithm performs a two-dimensional Fourier transform on the echo data in the frequency domain, achieving range and azimuth focusing through Stolt interpolation and phase factor compensation to obtain a three-dimensional complex image volume. The frequency domain imaging algorithm is approximately two orders of magnitude faster than the back projection algorithm, making it suitable for the throughput requirements of security inspection scenarios. After focusing, the three-dimensional image volume is projected with maximum intensity along the range dimension to obtain a two-dimensional complex SAR image. The pixel resolution is approximately 5mm in the range direction and approximately 5mm in the azimuth direction.

[0122] Simultaneously record imaging geometry The imaging geometry This includes the position of each array element, the horizontal scanning position sequence, the frequency sampling parameters, and the distance from the imaging plane to the array plane.

[0123] In this embodiment, the imaging area is The frontal projection area of ​​the human body is then divided into 12 overlapping patches, each patch being [size missing]. The overlap width between adjacent blocks is 80mm.

[0124] Step S2: Generate a basic shape physical response library based on the structural characteristics of the object to be detected and the imaging geometry.

[0125] For security inspection imaging scenarios, the structural characteristics of the objects to be inspected include the human body surface, clothing layers, and carried items. The basic shapes include the following types.

[0126] Point-like strong scattering bodies: Describes small metal objects such as metal buttons, zipper pulls, keys, or coins. Shape parameters Including the central location Equivalent scattering cross section and equivalent reflection coefficient .

[0127] Rectangular planar shape: Describes flat objects such as mobile phones, ID cards, knives, or cards. Shape parameters Including the central location ,length ,width Direction angle ,thickness and reflection coefficient .

[0128] Cylinder: Describes tubular objects (pens, tools, bottle-shaped containers, etc.). Shape parameters Including the central location ,length ,diameter Direction angle and reflection coefficient .

[0129] Irregular blocks: Describe large packages or containers. Shape parameters Including the central location Length of the circumscribed rectangle ,width Direction angle and equivalent reflection coefficient .

[0130] Curved surfaces of the human body: Describes the curved surfaces of the human torso, limbs, and other body parts. Shape parameters. Including the central location radius of curvature Arc angle and reflection coefficient .

[0131] Garment fold edges: Describes the edges and folds of garment layers. Shape parameters Including the central location ,length Direction angle and fold depth .

[0132] The response template is generated using a geometrical optics approximation method. This approximation utilizes physical optics (PO) surface element integration to calculate the scattered field distribution of each basic shape surface: the basic shape surface is discretized into surface elements, the induced current of each surface element is calculated based on the incident wave direction and the surface element normal, and then the far-field of the scattered data over all surface elements is integrated to obtain the scattering cross-section of the basic shape. This scattering cross-section is then substituted into the SAR focusing operator to generate a complex response template. For objects with metallic surfaces, the reflection coefficient is close to 1; for human skin and clothing, the reflection coefficient is calculated based on the material's dielectric constant.

[0133] In this embodiment, the response library is generated online on demand because the body shape and posture of the inspected object vary greatly in security inspection scenarios, and a fixed template library cannot cover all situations. The system estimates the approximate body shape and posture of the inspected object based on the initial low-resolution imaging results and dynamically generates a template family that matches the current scene. There are 100 templates for point-like strong scatterers, 1500 templates for rectangular planar objects (covering a length range of 50mm to 300mm and a width range of 20mm to 200mm), 800 templates for cylinders, 300 templates for irregular blocks, 500 templates for curved surfaces of the human body, and 800 templates for the edges of clothing folds. The total number of templates is approximately 4000.

[0134] Step S3: Detect candidate scattering features in the complex SAR image, match the candidate scattering features with the basic shape physical response library, and generate a candidate component set.

[0135] The detection method for candidate scattering features is the same as in Example 1. Since the human body surface itself generates a large area of ​​diffuse echo in security inspection scenarios, the threshold for detecting scattering peaks is adjusted to the local background mean plus 5 times the standard deviation. To avoid misdetecting normal human body echoes as candidate scattering features, the aspect ratio threshold for detecting thin, bright lines has been adjusted to 3:1 (the aspect ratio of items such as knives in security inspection scenarios may be smaller than that of cracks in industrial inspection).

[0136] The coarse matching method is the same as in Example 1. In the security inspection scenario, the system also utilizes human body region segmentation prior: first, the human body contour is extracted by low-frequency components, and the imaging area is divided into human body region and non-human body region, and candidate scattering feature detection is performed only in the human body region.

[0137] In this embodiment, step S3 detects approximately 150 candidate scattering features and generates the candidate component set containing 150 candidate components.

[0138] Step S4: Solve the candidate component set using an alternating iterative approach based on the complex field additive response model.

[0139] The additive response model in the complex domain is the same as in Example 1. In this example, This represents the number of pixels in a single tile (approximately 360,000). Initially 150. Noise energy estimation. (Noise levels in security inspection scenarios are higher than in industrial inspection scenarios), stability items Component number penalty coefficient Regularization weights .

[0140] In the contribution estimation step, the soft contribution weight is calculated in the same way as in Example 1.

[0141] In the parameter update step, this embodiment uses gradient descent to update the shape parameters, with a learning rate of... Complex amplitude parameters The update method is the same as in Example 1. During the update process, the constraint component parameters fall within the physically feasible range: the length of the rectangular planar body... cylinder diameter radius of curvature of the human body surface arc .

[0142] Because security check scenarios have high real-time requirements (single person passage time not exceeding 5 seconds), this embodiment adopts a hierarchical matching strategy to accelerate computation. The first layer is rapid candidate screening: using a low-resolution template (downsampling by 4 times), 20 rounds of alternating iterations are performed on all 150 candidate components to quickly eliminate obvious noisy candidates and human background components, reducing the number of components to approximately 50. The second layer is fine matching: using the original resolution template, alternating iterations are performed on the remaining 50 candidate components until convergence.

[0143] The rules for component management are the same as in Example 1. Pruning threshold. (In security inspection scenarios, the total contribution of small items may be small, so the pruning threshold should be appropriately reduced to avoid omissions.) The merging position distance threshold is 10mm, and the merging direction angle difference threshold is 10°.

[0144] In this embodiment, the first-layer fast screening takes about 1 second, and the second-layer fine matching converges after about 40 iterations, taking about 2 seconds. The 150 initial candidate components eventually converge to 45 components, including 18 curved human body surfaces, 10 clothing fold edges, 8 point-like strong scattering bodies, 5 rectangular planar bodies, 3 cylinders, and 1 irregular block.

[0145] Step S5: Determine the true support confidence, retain true structure pixels and suppress artifact pixels, and output the enhanced SAR image.

[0146] The calculation method for the true support confidence is the same as in Example 1, with the distance attenuation parameter... (In security inspection scenarios, both the size of the items and the imaging resolution are larger than in industrial inspection scenarios, so the distance attenuation parameter is adjusted accordingly.) Enhanced complex pixels and artifact residuals The calculation formula is the same as in Example 1.

[0147] In security check scenarios, the system categorizes curved surfaces of the human body and edges of clothing folds as background structures, attributing and suppressing their side lobes and ghost effects. Only the true structural responses corresponding to the carried items are retained in the enhanced image, significantly improving the contrast of the carried items relative to the background.

[0148] This embodiment also outputs contribution maps, artifact maps, and explanation information for defect candidate regions ("suspicious item candidate regions" in security inspection scenarios). The explanation information for each suspicious item candidate region includes: candidate region location, size, corresponding basic shape type (rectangular planar body, cylinder, irregular block, or point-like strong scattering body), true support confidence, and artifact confidence.

[0149] Step S6: Perform cross-tile consistency determination and weighted fusion on adjacent tiles.

[0150] The method for cross-tile consistency determination and weighted fusion is the same as in Example 1. The tile center distance attenuation parameter is used for fusion weights. For items carried across tiles, the system merges matching components in adjacent tiles into the same global item. The merging conditions are similar to those in Example 1: the positional distance is less than 20mm, the shape type is the same, and the directional angle difference is less than 15°.

[0151] In security imaging, traditional intensity map processing methods often confuse the sidelobe responses of human body surface folds and clothing edges with the true responses of small metal objects, resulting in a false alarm rate of approximately 25%. By using the method of this invention, the sidelobe and ghosting effects of the human body surface and clothing edges are modeled as the basic background structure, and their effects are attributed and suppressed. This improves the contrast of small metal objects in the enhanced image by approximately 6 dB and reduces the false alarm rate to 8%. A folding knife (120 mm long and 25 mm wide) hidden in a fold inside clothing is partially obscured by the sidelobe of the clothing fold in a traditional intensity map. This invention attributes the sidelobe of the clothing fold to the edge components of the clothing fold and marks them as artifacts, allowing the true support response of the folding knife to be fully revealed in the enhanced image. The total processing time for a single-person full-body scan is approximately 3.5 seconds (including 1 second for the first-layer rapid screening, 2 seconds for the second-layer fine matching, and 0.5 seconds for cross-tile fusion), meeting the security pass rate requirements.

[0152] In summary, the embodiments disclosed herein have at least the following technical effects: This invention achieves interpretable enhancement for large-scale millimeter-wave SAR imaging through steps including additive modeling in the complex domain, a physical response library for basic shapes, a soft contribution weighting mechanism in alternating iterative solutions, real support and artifact discrimination, and cross-tile consistency fusion. This method can improve the imaging clarity of complex objects while maintaining physical traceability, providing a more reliable data foundation for defect detection and manual verification. This invention can be integrated with existing millimeter-wave SAR acquisition equipment and imaging workflows, serving as a back-end enhancement module or extending to joint inversion of the original measurement domain, exhibiting good engineering compatibility.

[0153] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A millimeter-wave SAR imaging enhancement method, characterized in that, include: Step S1: Acquire millimeter-wave echo data, perform amplitude and phase calibration and focusing processing on the millimeter-wave echo data, generate a complex SAR image that retains the amplitude and phase of each pixel, and record the imaging geometry; Step S2: Based on the structural characteristics of the object to be detected and the imaging geometry, a basic shape physical response library is generated. The response template of each basic shape in the basic shape physical response library is a complex response generated by the millimeter wave propagation model and the SAR focusing operator. The response template includes a main lobe region marker and an artifact region marker. Step S3: Detect candidate scattering features in the complex SAR image, match the candidate scattering features with the basic shape physical response library to generate a candidate component set, wherein each candidate component in the candidate component set has an initial shape parameter and an initial amplitude and phase parameter; Step S4: Based on the complex domain additive response model, the candidate component set is solved by alternating iteration. In the contribution estimation step, the soft contribution weight of the pixel from different components is estimated according to the predicted complex contribution of each component at each pixel. In the parameter update step, the shape parameters and amplitude and phase parameters of each component are updated so that the complex responses of all components are superimposed to approximate the complex SAR image. The contribution estimation step and the parameter update step are executed alternately until convergence. Step S5: Determine the true support confidence of each pixel based on the soft contribution weight, the distance between the pixel and the physical support region of the corresponding component, the phase consistency between the component's predicted response and the observed pixel, and the artifact region marker in the response template. Based on the true support confidence, retain the true structure pixels and suppress the artifact pixels, and output the enhanced SAR image.

2. The millimeter-wave SAR imaging enhancement method according to claim 1, characterized in that, In step S2, the basic shapes include point-like strong scatterers, linear cracks, edges, holes, welds, and arc surfaces, each of which is defined by shape parameters. Description, the shape parameters The response template includes center position, size, orientation angle, and equivalent reflection coefficient; the generation method of the response template includes one or a combination of analytical approximation, geometrical optics approximation, point scattering discretization, electromagnetic simulation, or experimental calibration; the artifact region marker distinguishes each pixel in the response template into main lobe region, side lobe region, ghost region, or out-of-focus region.

3. The millimeter-wave SAR imaging enhancement method according to claim 1, characterized in that, In step S3, scattering peaks, thin bright lines, and regions with local curvature changes are detected in the amplitude map of the complex SAR image as candidate scattering features, and background noise is eliminated by utilizing the complex phase continuity; the candidate component set allows for overcomplete initialization, and the number of candidate components in the candidate component set is greater than the actual number of structures in the object to be inspected.

4. The millimeter-wave SAR imaging enhancement method according to claim 1, characterized in that, In step S4, the complex domain additive response model is expressed as: ； in, Let be the complex SAR image vector. Number of pixels For the total number of components, For the first Complex amplitude parameters of each component, For the first Each component in the imaging geometry Complex SAR response under the following conditions For the first The shape parameters of each component This includes noise and unmodeled components.

5. The millimeter-wave SAR imaging enhancement method according to claim 4, characterized in that, In the contribution estimation step, the first... The component in the first The prediction complex contribution at each pixel is The soft contribution weight The calculation formula is: ； in, The complex SAR response In the The component at each pixel For noise energy estimation, For stable terms; when the responses of multiple components overlap on the same pixel, multiple Both are greater than zero.

6. The millimeter-wave SAR imaging enhancement method according to claim 5, characterized in that, In the parameter update step, the shape parameter and amplitude parameter of each component are updated with the following objective function: ； in, For regularization terms, The weight coefficients of the regularization term; the regularization term The shape parameters are constrained to fall within a physically feasible range.

7. The millimeter-wave SAR imaging enhancement method according to claim 6, characterized in that, Step S4 further includes component management: if the total soft contribution of a component in the entire pixel range is lower than a preset contribution threshold or the magnitude of its amplitude and phase parameters is lower than the noise level, then the component is deleted; if the similarity of the shape parameters and responses of two components is higher than a preset similarity threshold, then the two components are merged and the parameters are re-estimated; if the residual region corresponding to a component presents multiple spatially separated peaks, then the component is split into multiple sub-components.

8. The millimeter-wave SAR imaging enhancement method according to claim 7, characterized in that, In step S5, the enhanced complex pixels and artifact residuals The calculation formula is: ； in, The complex SAR image vector The pixel value, For pixels Belongs to components The confidence level of the real support based on real physical support. For the first step S4 The component in the first The predicted complex contribution at each pixel; Preserve the true structural contribution. The artifacts that are identified.

9. The millimeter-wave SAR imaging enhancement method according to claim 1, characterized in that, Step S1 further includes dividing the imaging area into multiple patches with overlapping boundaries, the overlap width being greater than the main lobe width and the side lobe width; after step S5, step S6 is also included: performing cross-pattern consistency determination on the component parameters in the overlapping area of ​​adjacent patches, when two components in adjacent patches are close in position, have the same direction and are connected at their length endpoints in global coordinates, the two components are merged into the same global structure, and weighted fusion is performed on the enhanced SAR image and artifact residual in the overlapping area.

10. The millimeter-wave SAR imaging enhancement method according to any one of claims 1 to 9, characterized in that, In step S5, a contribution map and an artifact map are also output. The contribution map records the distribution of the soft contribution weight of each pixel on each component, and the artifact map records the location and source component of the suppressed artifact pixels. Step S5 also outputs explanation information for each defect candidate region. The explanation information includes the main contributing component of the candidate region, the true support confidence, the artifact confidence, and the phase consistency score.

Images