A tire x-ray image intrinsic feature self-calibration method
By using a self-calibration method based on the inherent features of tire X-ray images, the problems of inconsistent pixel equivalents and distortion errors in tire X-ray inspection equipment were solved, realizing online self-calibration, improving inspection efficiency and accuracy, and simplifying the calibration process.
Patent Information
- Application Number
- CN202611452485.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-17
AI Technical Summary
Existing tire X-ray inspection equipment suffers from problems such as inconsistent pixel equivalents in different regions during global calibration, low efficiency of offline calibration, lack of distortion compensation, and reliance on manual labor, which affect the accuracy of local dimension measurement and inspection efficiency.
A self-calibration method based on the inherent features of tire X-ray images is adopted. By extracting the inherent structural features of the tire itself and combining them with a zonal calibration strategy, the actual pixel equivalent of each region is calculated to achieve online self-calibration. This includes offline global pre-calibration, online real-time dimensional calibration, and online self-updating calibration steps. The method utilizes the tire's inherent reference features to perform closed-loop matching with the true value of CAD design, and adaptively fine-tunes the pixel physical equivalent.
It simplifies the calibration process, reduces labor intensity, shortens calibration time, obtains more detailed calibration data, improves the accuracy and stability of dimensional measurement, and solves the problems of inconsistent pixel equivalents and distortion errors in different regions in traditional global calibration.
Smart Images

Figure CN123118786A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of tire X-ray inspection technology, and particularly relates to a method for self-calibration of inherent features of tire X-ray images. Background Technology
[0002] Currently, tire X-ray inspection equipment generally uses a global calibration method. This method is based on a set of metal calibration blocks and its accompanying image processing algorithm. During operation, multiple metal calibration blocks need to be manually attached to the tire surface, and the calibration data is obtained by reading the images through X-ray scanning.
[0003] However, the above calibration scheme has the following shortcomings: (1) Using multiple metal calibration blocks for global calibration does not take into account the cone beam projection amplification gradient error caused by the radial thickness difference of the tire crown, shoulder and sidewall areas, resulting in inconsistent pixel equivalents in each area and affecting the accuracy of local dimension measurement. (2) The calibration process is offline and requires periodic shutdowns to insert standard blocks for calibration, which causes production line interruptions and reduces testing efficiency; (3) Lacking a distortion compensation mechanism, it is unable to correct the geometric deviations introduced by the pincushion distortion at the detector edge and the focus drift of the X source; (4) The recognition of calibration block images relies on manual labor and has a low degree of automation. Summary of the Invention
[0004] This invention addresses the technical problems of inconsistent pixel equivalents in different regions, low efficiency of offline calibration, lack of distortion compensation, and reliance on manual labor in the global calibration method used in the aforementioned tire X-ray inspection equipment. It proposes a self-calibration method based on the inherent features of tire X-ray images. This method eliminates the need for separate metal calibration blocks; it directly extracts the inherent structural features and design dimensions of the tire itself as calibration references. Using a computer and corresponding software algorithms, calibration data is automatically calculated based on the difference between the actual imaging position and the theoretical design position of these features in the X-ray image, thus achieving online self-calibration.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is: a method for self-calibration of inherent features of tire X-ray images, comprising: Offline global pre-calibration steps: Obtain the pixel physical equivalents and distortion matrices corresponding to the five regions of the tire as initial calibration parameters; Online real-time size calibration steps: After distortion correction, physical equivalent conversion of pixels in each region, and dual-projection symmetry correction, the detection results are output; Online self-updating calibration steps: Trigger the calibration process, construct and solve the global distortion fitting equation based on the inherent reference characteristics of the tire and the theoretical dimensions of the tire design, obtain the updated calibration parameters and update them online, and achieve parameter locking or iterative optimization through convergence judgment.
[0006] In one embodiment, in the offline global pre-calibration step, the five tire regions are divided according to the radial structure of the tire by geometric or semantic segmentation methods. The five tire regions include the left sidewall, left shoulder, crown, right shoulder, and right sidewall in sequence along the tread.
[0007] In one embodiment, the online real-time size calibration step includes the dual-projection symmetry correction step, which involves: acquiring X-ray images of the tire in two rotational poses at 0° and 180°, measuring the pixel equivalent of each inherent reference feature of the tire in the 0° and 180° poses, and calculating the symmetry mean.
[0008] In one embodiment, in the online self-updating calibration step, the triggering condition for triggering the calibration process is at least one of timed triggering, tire specification change triggering, or multiple consecutive sets of dimensional deviations exceeding the threshold triggering; the inherent reference features of the tire include the bead wire bundle, the belt layer parallel wires, and the sidewall carcass cords.
[0009] In one embodiment, the inherent reference features of the tire are obtained using the following method: Denoising and grayscale normalization preprocessing were performed on tire X-ray images; Each inherent reference feature is extracted separately, and the inherent reference features of the tire are separated from the background; Extract the centerline of the inherent reference feature and perform sub-pixel level fine localization of the edge of the inherent reference feature; Geometric fitting yields measured data of the tire's inherent reference characteristics.
[0010] In one embodiment, constructing the global distortion fitting equation includes the following steps: The distortion model, cone-beam magnification model, and calibration constraints are combined and incorporated into a projection model described by parameter vector θ, which is defined as follows: , In the formula, Introduced only as symbols for unknown quantities; SDD represents the distance from the X-ray source to the detector, an inherent constant of the equipment; p_d represents the inherent pixel pitch of the detector, in mm / pixel, an inherent constant of the equipment; SOD1 represents the distance from the X-ray source to the left side of the tire; SOD2 represents the distance from the X-ray source to the left shoulder of the tire; SOD3 represents the distance from the X-ray source to the tire crown; SOD4 represents the distance from the X-ray source to the right shoulder of the tire; SOD5 represents the distance from the X-ray source to the right side of the tire. x represents the horizontal offset of the pixel coordinate; y represents the vertical offset of the pixel coordinate; k1, k2, and k3 are the radial distortion coefficients; p1 and p2 are the tangential distortion coefficients. The design theoretical dimension coordinates of the inherent reference features of the tire are projected forward through the projection model to obtain the predicted image points, which are then compared with the measured image points to construct the residual vector and the global cost function. When the number of features is greater than the number of unknowns, an overdetermined fitting equation is formed.
[0011] In one embodiment, a distortion model is used to correct detector edge distortion, employing the Brown-Conrady model, with the following formula: , , In the formula, x_d represents the normalized abscissa of the distorted image plane; x represents the normalized abscissa of the image plane; r represents the radial distance from the normalized image plane coordinate to the principal point; and y represents the normalized ordinate of the image plane. y_d represents the normalized ordinate of the image plane after distortion.
[0012] In one embodiment, the cone-beam amplification model is M_i = SDD / SOD_i, s_i=p_d / M_i=p_d·SOD_i / SDD, In the formula, M_i represents the geometric magnification; SOD_i represents the distance from the ray source to the plane of the object in the i-th region, i = left tire sidewall, left tire shoulder, tire crown, right tire shoulder, right tire sidewall; s_i represents the physical equivalent of the pixel in the i-th region, in mm / pixel; p_d represents the detector's inherent pixel pitch, in mm / pixel, which is an inherent constant of the device.
[0013] In one embodiment, the calibration constraint formula is: s_i'=L_{cad,k} / N_k, In the formula, s_i' represents the measured value of the physical equivalent of the pixel in the i-th region, in mm / pixel; L_{cad,k} represents the theoretical design size of the k-th inherent reference feature in the i-th region; and N_k represents the measured number of pixels on the image.
[0014] In one embodiment, the convergence determination includes: The tire was retested using the updated calibration parameters, and the deviation between the measured values and the theoretical tire design dimensions was calculated. If the deviation is less than the preset tolerance or the relative rate of change of the deviations of two adjacent rounds is less than the preset threshold, then the deviation is determined to have converged, self-calibration is completed, and the parameters are locked. Otherwise, the retest results will be used as new calibration constraints to enter the next iteration until the convergence condition is met or the maximum number of iterations is reached, at which point the system will be forcibly locked.
[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The self-calibration method of inherent features of tire X-ray images of the present invention simplifies the existing calibration process, reduces labor intensity, and shortens calibration time; at the same time, since no metal block is used, more detailed calibration data can be obtained; through independent calibration of partitions, dual-projection symmetrical correction and distortion compensation, the problems of inconsistent pixel equivalents and distortion errors in different regions of traditional global calibration are solved, which is conducive to improving the accuracy and stability of subsequent dimensional measurements. Attached Figure Description
[0016] Figure 1 This is an overall flowchart of the tire X-ray image inherent feature self-calibration method of the present invention; Figure 2 A schematic diagram showing the arrangement of a multi-zone arc-shaped target and a rotating mechanism; Figure 3 A schematic diagram of the structure of a multi-zone arc-shaped target; Figure 4 Flowchart for offline global pre-calibration; Figure 5 This is a schematic diagram of five-region segmentation in a tire X-ray image; Figure 6 Flowchart for online real-time dimensional calibration; Figure 7 Flowchart for online self-updating calibration; In the diagram: 1. Multi-zone arc-shaped target; 11. Mounting base; 12. Telescopic drive mechanism; 121. Telescopic guide rail; 122. Power assembly; 13. Arc-shaped contour frame; 131. Left tire sidewall arc-shaped section; 132. Left tire shoulder arc-shaped section; 133. Tire crown arc-shaped section; 134. Right tire shoulder arc-shaped section; 135. Right tire sidewall arc-shaped section; 14. Standard metal block; 2. Rotating mechanism; 3. Encoder; 4. X-ray source; 5. Flat panel detector; 6. PLC controller; 71. Left side of the fetus; 72. Left shoulder of the fetus; 73. Fetal crown; 74. Right shoulder of the fetus; 75. Right side of the fetus. Detailed Implementation
[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Because the tire is a rotating structure, the radial thickness of the tire crown 73, shoulder, and sidewall regions varies, resulting in differences in the surface area distance (SOD) from the X-ray source 4 to each region. According to cone-beam projection geometry, different object distances correspond to different magnifications, M = SOD / SOD, leading to inconsistent pixel equivalents in different regions on the same detector. Existing global calibration methods use metal calibration blocks for uniform global calibration, ignoring the projection magnification differences caused by the aforementioned radial thickness gradient variations along the tire.
[0019] To address the aforementioned problems, this invention provides a method for self-calibrating the inherent features of tire X-ray images. By extracting the inherent structural features of the tire itself and combining them with a zonal calibration strategy, the actual pixel equivalent of each region is calculated, achieving zonal self-calibration. The self-calibration refers to: without the aid of external metal calibration blocks, using the tire's inherent reference features—the metal skeleton features that are always clearly imaged in the X-ray image and whose geometric dimensions are strictly defined by the CAD design documents—as the calibration reference; based on the ratio between the measured pixel size on the image and the theoretical physical size in the CAD document, the geometric and distortion parameters of the imaging system are inferred, thereby obtaining the pixel physical equivalent of each imaging region, in mm / pixel.
[0020] refer to Figure 1 As shown, the tire X-ray image inherent feature self-calibration method of the present invention includes S1, offline global pre-calibration step; S2, online real-time size calibration step; and S3, online self-updating calibration step. The specific process is as follows: Upon startup, a self-test is performed to determine if monthly offline calibration is required. If required, offline full-domain pre-calibration is performed: a multi-zone arc-shaped target is automatically extended while the equipment is stopped; 360° rotation imaging is performed; distortion coefficients are calculated; regional cone-beam magnification is fitted; calibration parameters are stored; the target is retracted; and then online detection begins. If not required, online real-time dimensional calibration is performed while the production line is running continuously without stopping. Tires are loaded, and rotation angles and tire specifications are simultaneously collected. X-ray imaging is performed, and the image is segmented into five regions. The rubber thickness of each region is obtained, and the corresponding pixel physical equivalent and distortion correction matrix are retrieved. The pixel dimensions are converted to the actual physical dimensions, and the dimensional detection results are output after 0° and 180° dual-projection symmetry error compensation. The system then determines whether online self-updating calibration is triggered. If triggered, it is performed while the production line is running continuously without stopping, without the need for an external target. The inherent reference characteristics of the tire are used, and a closed-loop matching with the true value of the CAD design is performed. The pixel physical equivalent of each region is adaptively fine-tuned, calibration parameters are iteratively optimized, and the database is updated. Otherwise, the system continuously detects the next tire.
[0021] It should be noted that the above three steps of the present invention share the same closed-loop framework of "feature extraction → comparison with the ground truth → obtaining or updating pixel physical equivalent and distortion parameters", as well as the same five-region segmentation and distortion correction matrix, but differ in calibration benchmark, triggering conditions and whether to stop the machine.
[0022] The offline full-domain pre-calibration step of this invention is performed after monthly or major adjustments, while the equipment is shut down. It uses an external multi-zone arc-shaped target 1 to obtain the initial values of the full-width distortion and the initial physical equivalent of the pixels in the five regions of the flat panel detector 5, providing initial calibration parameters for subsequent online detection. The working principle is as follows: using the known metal cube dimensions of the multi-zone arc-shaped target 1 as the true value benchmark, an overdetermined fitting equation is established based on the distortion model and the cone-beam amplification model. Robust least squares iterative solution is then used to obtain the initial calibration parameters for the equipment.
[0023] The offline global pre-calibration step of this invention adopts Figure 2 The multi-zone arc-shaped target 1 and the rotating mechanism 2 are shown. The design principle of the multi-zone arc-shaped target 1 is as follows: Since the tire is a curved rotating body, the object distance from each region to the X-ray source 4 is different. The target adopts an arc-shaped multi-zone structure to ensure that the standard features of each region are at the same object distance as the regions of the real tire, thereby establishing the thickness-magnification mapping relationship and the pixel physical equivalent lookup table respectively. If a planar target is used, the object distance difference between the tire crown 73 and the tire sidewall will be ignored, and the pixel physical equivalent of the partition will be inaccurate. The object distance from the location of the standard feature in each region to the X-ray source 4 is equal to the nominal object distance SOD_i of the corresponding region of the real tire. The object distance of the tire crown 73 is the smallest, and the object distance of the tire sidewall is the largest, thereby ensuring that the magnification of each region is consistent with that of the real tire. At least one metal feature of known standard size is arranged on the arc surface of each arc segment. The metal feature can be a standard square metal sheet, a circular hole array, or a grid, and the material is a high-absorbent metal for X-rays such as lead or tungsten steel. The processing size of the metal feature is used as the true value for pixel physical equivalent calibration.
[0024] The multi-zone arc-shaped target 1 is an arc-shaped calibration fixture used only during offline global pre-calibration, driven in and out of the ray optical path by a telescopic drive mechanism 12. (Reference) Figure 3As shown, the multi-zone arc-shaped target 1 includes: a mounting base 11, disposed at the imaging station; a telescopic drive mechanism 12, disposed on the mounting base 11, including a telescopic guide rail 121 and a power component 122. The telescopic guide rail 121 is disposed on the mounting base 11 and supports the arc-shaped contour frame 13. The power component 122 is disposed on the mounting base 11 and includes a cylinder or a servo electric push rod for driving the arc-shaped contour frame 13 to extend or retract the multi-zone arc-shaped target 1; the arc-shaped contour frame 13 is slidably mounted on the telescopic guide rail 121. The target 13 consists of five arc-shaped segments arranged sequentially along the arc surface, including the left tire sidewall arc segment 131, the left tire shoulder arc segment 132, the tire crown arc segment 133, the right tire shoulder arc segment 134, and the right tire sidewall arc segment 135, corresponding to the left tire sidewall 71, left tire shoulder 72, tire crown 73, right tire shoulder 74, and right tire sidewall 75 in the five tire regions, respectively. The radius of curvature of each segment is determined according to the nominal contour radius of the corresponding tire region, so that the target arc surface is parallel to the outer contour of the tire in the axial section. A standard metal block 14 of known size is set on the arc surface of each arc segment. During calibration, the power component 122 drives the arc-shaped contour frame 13 to extend the multi-zone arc-shaped target 1 into the X-ray optical path. After calibration, the power component 122 drives the arc-shaped contour frame 13 to retract the multi-zone arc-shaped target 1 out of the optical path to avoid obstructing the online detection.
[0025] The rotating mechanism 2 is installed at the X-ray machine imaging station, located on the tire clamping or conveying mechanism, and is controlled by the PLC controller 6. The rotating mechanism 2 is used to drive the multi-zone arc target 1 or the tire to rotate at a constant speed around the tire rotation axis. The encoder 3 is coaxial with the rotating mechanism 2 and synchronously records the rotation angle, providing an angle reference for 0° and 180° dual projection and multi-frame image acquisition.
[0026] refer to Figure 4 As shown, the offline global pre-calibration steps specifically include: S11. After monthly calibration or major system adjustments, the equipment PLC controller 6 triggers an automatic calibration command; S12. The multi-zone arc-shaped target 1 extends into the X-ray optical path and enters the X-ray imaging field of view. At this time, there is no tire obstruction in the optical path. S13. The rotating mechanism 2 drives the multi-zone arc target 1 to rotate uniformly around the axis of rotation 360°. The encoder 3 records the angle at fixed angular intervals and acquires X-ray projection images in multiple frames to provide over-constrained data for solving distortion and magnification gradients. Preferably, the fixed angular interval is 15°. S14. Perform image processing, sequentially processing the multi-frame X-ray projection images as follows to extract the corner pixel coordinates of the standard metal cube 14 in each region at the sub-pixel level: S141. Preprocessing: Anisotropic diffusion denoising, grayscale normalization and local contrast enhancement; S142, ROI coarse localization: Based on the known positions of the five target regions in the image and the encoder 3 angle information, the regions of interest where the standard metal cubes 14 in each region are located are extracted; S143, Block segmentation: Extract the binary connected regions of blocks within the ROI using adaptive thresholding Otsu combined with morphological opening and closing operations or Canny edge detection. S144, Subpixel Edge Extraction: The Zernike moment method or subpixel Canny is used to accurately locate the edge of each square; S145. Corner point calculation: Perform least squares or Hough fitting on each edge line, and obtain the sub-pixel coordinates of the four corner points of the block by taking the intersection of adjacent edges. ; S146, Multi-angle fusion: Utilize multi-frame data with the target rotated 360° and each frame being 15° apart, perform least-squares averaging or RANSAC to remove outlier frames for the same corner point of each block in each frame, and output the pixel coordinates of the 14 corner points of a high-confidence standard metal block.
[0027] S15. Using the physical dimensions of the standard metal block 14 as the true value, solve for the full-width radial distortion correction coefficients k1, k2, k3 and tangential distortion correction coefficients p1, p2 of the flat panel detector 5, and generate the distortion correction mesh. S16. Based on the cone-beam geometry formula M_i=SDD / SOD_i, five sets of thickness-magnification mapping relationships for left tire side 71, left tire shoulder 72, tire crown 73, right tire shoulder 74, and right tire side 75 are calculated from the known object distance of each region, and a lookup table for the pixel physical equivalent s_i of each region is established. S17. Store the partition calibration parameters and distortion matrix in the equipment database for online real-time dimension calibration and online self-updating calibration processes to call them online. S18, Multi-zone arc-shaped target 1 retracted, switch to online detection mode.
[0028] In the offline global pre-calibration step of this invention, five regions are segmented and referenced. Figure 5 As shown, the five imaging sub-regions obtained by dividing a single-frame tire X-ray fluoroscopic image according to its radial structure are, in order along the lateral side of the tire tread: left sidewall 71, left shoulder 72, tire crown 73, right shoulder 74, and right sidewall 75.
[0029] Because the tire is a rotating body, the radial thickness of the rubber varies in different areas: the crown (73) is thicker, resulting in a smaller object distance to the ray source (4), a larger cone-beam projection magnification M=SDD / SOD, and a smaller corresponding pixel physical equivalent (in mm / pixel); the sidewall is thinner, resulting in a larger object distance, a smaller magnification, and a larger corresponding pixel physical equivalent. If a uniform pixel physical equivalent is used without different zones, systematic deviations will occur in dimensional measurements. This invention is specifically designed to address this cone-beam projection magnification gradient error.
[0030] The five-region segmentation method adopts a geometric method or a semantic segmentation method, and the two methods can be implemented in parallel. The geometric method first uses edge detection combined with circle or ellipse fitting to extract the inner and outer contours of the tire to obtain the symmetry axis of the tread 73; then performs sampling at equal intervals along the axial direction, according to the formula t(x)=y_out(x) y_in(x) to calculate the local radial thickness of each sampling point, and the partition boundary is determined according to the thickness ratio method or the thickness threshold method. For the thickness ratio method, taking the maximum thickness t_max of the tread 73 as a reference, proportional coefficients α1 and α2 are set, for example, α1=0.75 and α2=0.40 are set, the tread 73 region satisfies t / t_max≥α1, the shoulder region satisfies α2≤t / t_max<α1, and the sidewall region satisfies t / t_max<α2. For the thickness threshold method, absolute thresholds T1 and T2 are set according to the CAD design thickness of the tire specification, for example, T1=12mm and T2=7mm are set, the tread 73 region satisfies t≥T1, the shoulder region satisfies T2≤t<T1, and the sidewall region satisfies t<T2. Radial dividing lines are drawn at the boundary positions x=±x1, ±x2, such as Figure 5 shown by the middle dashed line, and left-right symmetry obtains five sub-regions: left sidewall 71, left shoulder 72, tread 73, right shoulder 74, and right sidewall 75. Then the radial thickness t_i of the rubber in each region is counted for the subsequent independent correction of the pixel equivalent s_i and the distortion matrix.
[0031] The semantic segmentation method uses an encoder-decoder structured fully convolutional neural network (such as UNet, DeepLabv3+) to perform pixel-level semantic segmentation on the perspective image after distortion correction, classifying each pixel into one of five categories: left sidewall 71, left shoulder 72, tread 73, right shoulder 74, and right sidewall 75, and directly outputting a five-category region mask with the same resolution as the input. The specific steps are as follows: Data preparation: Collect X-ray perspective images of tires of multiple specifications, generate five types of pixel-level ground truth masks through expert annotation or automatically with the aid of the geometric method, and construct a labeled training set; Preprocessing: Perform distortion correction on the input image using the distortion matrix calibrated offline in the full domain, and perform grayscale normalization and denoising processing; Network inference: Feed the preprocessed image into the trained model, output a five-channel category probability map, and obtain the category label of each pixel through softmax and argmax, that is, the five-category region mask; Post-processing: Perform connected domain analysis, morphological smoothing and fine fragment region removal on the mask to ensure that each region is a single connected region; Thickness statistics: For each category i, extract the inner and outer boundaries of the region along the axial direction, calculate the local radial thickness and take the average value to obtain the rubber radial thickness t_i of the region; Equivalent assignment: Based on t_i, retrieve the physical equivalent s_i of the pixels in this area from the database, and superimpose s_i and the distortion matrix onto this area to complete the pixel-level geometric correction of the partition.
[0032] Semantic segmentation and geometric methods are parallel solutions and can be used independently; alternatively, the geometric method can be used to generate pseudo-labels to train the network first, and then the semantic segmentation method can be used for inference online, with the two methods mutually verifying each other.
[0033] This invention's online real-time dimensional calibration step is executed during continuous production line operation without downtime. It directly calls pre-stored calibration parameters to complete the dimensional measurement and defect classification of each tire, without performing parameter solving itself. Figure 6 As shown, it includes the following steps: S21. The tire to be inspected is conveyed to the X-ray imaging station via a conveyor belt, and the photoelectric sensor triggers the imaging; the encoder 3 simultaneously collects the tire rotation angle and tire model, and retrieves the standard CAD dimension library according to the model. S22, X-ray flat panel detector 5 acquires complete transparent images of the tire; S23. Divide the image into five imaging regions using semantic segmentation or geometric contouring methods, and distinguish the radial thickness t_i of the rubber in each region; S24. Retrieve the physical equivalent parameters s_i of the partition pixels from the offline global pre-calibration and apply distortion correction: S241, Partition parameter retrieval: Based on the five-region segmentation results in step 3, determine the region to which each pixel of the image belongs, retrieve the pixel physical equivalent s_i and corresponding distortion parameters corresponding to the region from the database obtained by offline global pre-calibration, where the physical equivalent of the 73 region of the tread is the smallest and the physical equivalent of the tread side region is the largest. S242, Distortion Correction: Using the distortion coefficients k1, k2, k3, p1, p2 obtained in the offline global pre-calibration step S15, or the distortion correction mesh generated therefrom, the original image is remapped to restore the pixels from the distorted positions to the ideal distortion-free positions. S243, Zone Equivalent Conversion: The corrected image is converted according to the physical equivalent of pixels s_i for each region. The physical length L = pixel length N × s_i. Since the magnification M_i of each region is different, the conversion must be done independently for each region and cannot be uniformly calculated globally. S244. Output: Obtain the distorted and pixel equivalent field corrected image, along with the pixel physical equivalent s_i of each region and the distortion matrix, for subsequent steps to perform feature size conversion. S25. Convert the pixel length to physical size for features such as bubbles, steel wires, and belt layers to obtain preliminary physical dimensions; S26. Obtain X-ray images of the same tire in two rotational poses at 0° and 180°. For the bead wire bundle, the parallel wires of the belt layer, and the tire sidewall cord, after imaging at two rotational poses of 0° and 180°, the pixel equivalents corresponding to each inherent reference feature are measured and denoted as m_feature(0°) and m_feature(180°), respectively. The symmetrical mean of each inherent reference feature is calculated as: m_feature=[m_feature(0°)+m_feature(180°)] / 2, thus forming a symmetrical measurement pair of the same feature for each type of feature, eliminating the focus drift error of the X-ray source 4. Correction principle: The drift Δ of the X-ray source 4 focus relative to the detector center has opposite signs at 0° and 180° rotation poses. After symmetrical averaging, it is canceled out, thereby eliminating the linear error introduced by the X-ray source 4 focus offset and the slight eccentricity of the workpiece. S27. Output the actual dimensions of defects and the measurement data of wire spacing, and generate quality inspection records.
[0034] This invention's online self-updating calibration step is performed during continuous production line operation without downtime. It requires no external target and utilizes the tire's inherent baseline features and a closed loop with the CAD design truth value to adaptively fine-tune the pixel equivalent of each region, i.e., model self-calibration. Figure 7 As shown, it includes the following steps: S31. The system automatically updates itself when triggered by timed or abnormal conditions. The triggering conditions are at least one of timed triggering, changing tire specifications, or multiple consecutive sets of dimensional deviations exceeding the standard. S32. When the triggering condition is met, based on the image automatic extraction algorithm link, three types of inherent reference features are located on the current tire X-ray fluoroscopic image, including the bead wire bundle, the parallel wires of the belt layer, and the sidewall carcass cords. Among them, the bead wire bundle: the annular wire loops at both ends of the tire appear as high grayscale closed loops in the X-ray image. The measured bead diameter D_bead_img is obtained by fitting the inner and outer edges, with the unit being pixels; the parallel wires of the belt layer in the 73rd tread: the belt layer wires in the 73rd tread area appear as multiple approximately parallel high grayscale thin lines in the X-ray image. The measured wire spacing d_img is obtained by extracting the center lines of adjacent wires, with the unit being pixels; the carcass cords in the sidewall area: the sidewall is made of rubber, but the carcass cords that run through the sidewall can be clearly imaged in the X-ray fluoroscopic image. These cords run radially through the entire sidewall from the bead and extend to below the 73rd tread, appearing as a set of thin, high grayscale straight lines arranged axially and extending radially in the sidewall area. The ROI of the sidewall region is extracted, and after denoising, adaptive threshold segmentation and morphological refinement, the radial centerline of each cord is fitted by least squares or Hough to obtain the radial projection span L_img of each cord in the sidewall region and the axial spacing g_img of adjacent cords, in pixels. The length of the exposed cord section L_cad or the axial spacing g_cad designed by CAD is used as the true value to back-calculate the physical equivalent of the sidewall region pixels, which is used for cross-validation of the calibration results of this region. The three types of inherent benchmark features described in this invention share the same automatic image extraction algorithm link, including: Image preprocessing: Anisotropic diffusion or nonlocal mean denoising is performed on the X-ray perspective image to suppress speckle noise, followed by grayscale normalization to provide clean input for subsequent segmentation; ROI selection: Based on the five-region segmentation results, the bead region, the 73 belt layer region of the crown, and the sidewall region are selected respectively to reduce the processing range and increase the speed. Adaptive threshold segmentation: Local adaptive thresholding (Gaussian adaptive / Otsu local) is used to binarize and separate high-grayscale metallic features from the rubber background; Morphological denoising and thinning: Opening operations are used to remove small noise connected regions remaining after segmentation, skeleton thinning is used to obtain a single-pixel wide feature center line, and closing operations are used to bridge the gaps and ensure feature continuity. Subpixel edge fine localization: First, use Canny to extract pixel-level edges of the bead closure ring and the steel wire, and then use the Zernike moment method to perform subpixel-level fine localization along the edge normal direction (the localization accuracy is better than the pixel level) to obtain a high-precision subpixel edge point set; Geometric fitting: Geometric fitting is performed according to feature type. The bead wire bundle is fitted with circular or elliptical least squares fitting to fit the sub-pixel edge point set to obtain the inner and outer edges of the bead and the measured bead diameter. The parallel wires of the belt layer and the sidewall cords are fitted with the skeleton centerline using linear least squares fitting or Hough transform fitting to obtain the linear parameters. Then, the measured values of the spacing between adjacent wires of the belt layer or the radial span and axial spacing of the sidewall cords are calculated. S33. Import CAD physical true values: Based on the tire specification barcode, retrieve the corresponding specification's design theoretical values from the standard CAD dimension library, including the design spacing d_cad between adjacent steel wires in the belt layer, the design width W_belt of the belt layer, the standard diameter of the bead D_cad, and the radial design thickness t_i of the rubber in each area (i=left sidewall 71, left shoulder 72, crown 73, right shoulder 74, right sidewall 75). S34. Construct the global distortion fitting equation By combining the distortion model, the cone-beam magnification model and the calibration constraints, a global overdetermined fitting equation is established with the imaging geometric parameters as unknowns. S341. The distortion model adopts the Brown-Conrady model to correct the edge distortion of the flat panel detector 5. The distortion type is mainly pincushion distortion. Let the normalized coordinates of the 5-pixel coordinates (u, v) of the flat panel detector relative to the principal point (u0, v0) be: , , In the formula, x and y are normalized image plane coordinates, which are derived from pixel coordinates offset by the principal point. Obtained by translation; r is the radial distance from the image point to the principal point; The Brown–Conrady model is: , , In the formula, x_d represents the normalized abscissa of the image plane after distortion; k1, k2, and k3 are radial distortion coefficients (dimensionless), which originate from the electro-optical distortion and edge vignetting of the flat panel detector 5; p1 and p2 are tangential distortion coefficients (dimensionless), which originate from the non-ideal alignment of the flat panel detector 5 with the optical axis and the eccentricity of the panel assembly; the above five distortion coefficients are all unknowns to be solved by calibration iteration; y_d represents the normalized ordinate of the image plane after distortion. The derivation is based on the fact that the edge image points of the flat panel detector 5 have radial and tangential displacements relative to the ideal position. In engineering, the Brown–Conrady model formula is obtained by approximating the image points with radial polynomials (k terms) and tangential bilinear terms (p terms). S342, Cone-beam amplification model M_i = SDD / SOD_i, In the formula, M_i represents the geometric magnification; SDD represents the distance from the X-ray source 4 to the detector, which is an inherent constant of the device; SOD_i represents the distance from the X-ray source 4 to the plane of the object in the i-th region, i = left tire side 71, left tire shoulder 72, tire crown 73, right tire shoulder 74, right tire side 75; s_i=p_d / M_i=p_d·SOD_i / SDD, In the formula, s_i represents the model value of the physical equivalent of the pixel in the i-th region, in mm / pixel; p_d represents the inherent pixel pitch of the detector, in mm / pixel, which is an inherent constant of the device; SOD_i varies with the radial design thickness t_i of the rubber in the corresponding region, that is, a larger thickness in the tread 73 region results in a smaller SOD_i, a larger geometric magnification M_i, and a smaller pixel physical equivalent s_i, while a smaller thickness in the sidewall region results in a larger SOD_i, a smaller magnification M_i, and a larger pixel physical equivalent s_i; SOD_i is an unknown quantity to be determined or corrected, obtained by calibration inversion; S343, Calibration Constraints s_i'=L_{cad,k} / N_k, In the formula, s_i' represents the measured value of the physical equivalent of the pixel in the i-th region; L_{cad,k} represents the theoretical tire design size of the k-th inherent reference feature in the i-th region; N_k represents the number of measured pixels on the image; It should be noted that step S342 gives the geometric formula for the model value of the physical equivalent of the pixel in the i-th region, and step S343 gives the definition and measurement formula for the measured value of the physical equivalent of the pixel in the i-th region. The two are equivalent to each other: For an inherent reference feature located in the object plane of the i-th region with a tire design theoretical size of L_{cad,k}, its image length on the detector after cone-beam projection is L_{cad,k}·M_i. The number of pixels calculated according to the detector pixel pitch p_d is N_k=L_{cad,k}·M_i / p_d. Therefore, s_i'=L_{cad,k} / N_k=p_d / M_i=p_d·SOD_i / SDD, That is, the two equations are strictly equivalent; since SOD_i is an unknown quantity, setting the two equations equal yields the result. SOD_i=(L_{cad, k} / N_k)·(SDD / p_d), This is the constraint equation formed by combining the cone-beam magnification model and the calibration constraints for the unknown quantity SOD_i. In actual imaging, N_k is also affected by the distortion coefficients k1, k2, k3, p1, p2 and the principal point offset. Therefore, in step S344, a complete projection model with distortion is used to describe it uniformly. Multiple inherent reference features, five regions and 0° and 180° dual projections provide equations that are far more numerous than the number of unknown quantities, forming an overdetermined fitting equation and solving it uniformly.
[0035] S344. Establishment of Overdetermined Fitting Equations The distortion model, cone-beam amplification model, and calibration constraints are unified into a single projection model described by the parameter vector θ, allowing the same set of θ to simultaneously interpret the three types of equations, thus forming a simultaneous constraint on θ. The specific method of this simultaneous constraint is as follows: The CAD design physical coordinates (X_k, Y_k) of the k-th inherent reference feature are projected onto the model to obtain the predicted image point. : , The parameter vector θ is defined as follows: , Here It is introduced only as a symbol for unknown quantity to describe the composition of the undetermined parameters on which the projection model depends. Its specific value is unknown and will be determined by iterative solution in step S35. The residual vector is obtained by comparing it with the measured image points (u_k, v_k) on the image. ; Constructing a global cost function , In the formula, ρ represents the robust cost (e.g., Huber), which corresponds to the least squares after the RANSAC coarse removal in step 35.
[0036] Since each inherent baseline feature provides two residual equations x and y, while there are 12 unknowns θ to be determined (only 5 SODs, 2 offsets, and 5 distortion coefficients when SDD and p_d are fixed as constants); when the actual number of extracted features (three types of inherent features, multiple sampling points, and 0° and 180° dual projections) is much greater than the number of unknowns, the number of equations is much greater than the number of unknowns, thus forming an overdetermined fitting equation.
[0037] S35. Robust least squares iterative solution of imaging geometric parameters The overdetermined fitting equation is solved by combining RANSAC with Levenberg-Marquardt minimization of E(θ), outputting the pixel equivalents s_i of the five regions and the distortion correction matrix; specifically, based on the residual equation system r_k(θ) and the global cost function E(θ) established in step S344, the optimal solution θ of the parameter vector θ that minimizes E(θ) is found. That is, to solve for the specific value of the unknown θ introduced in symbolic form in step S344. The initial value of θ is given in step S351, which includes the following steps: S351, Initialization: Take the initial full-width distortion value and the initial SOD value of the object distance in the five regions obtained from offline pre-calibration. 0 _i and the initial value of detector offset (Δx) 0 Δy 0 ) is used as the initial value θ0 for the Levenberg–Marquardt iteration; S352, RANSAC Robust Coarse Removal: For the k-th inherent datum feature, its CAD design physical coordinates (X_k, Y_k) are projected onto the model using the current model. Forward projection yields the predicted image points. Solve the reprojection residuals using the measured image points (u_k, v_k) of the image. (pixels); Set anomaly criteria: Obtain subpixel localization noise σ (approximately 0.1–0.3px after Zernike fine localization) from the detector background calibration, and take a threshold T=3σ (approximately 0.5–1.0px); or use T=c·MAD(r_1…r_K), where c takes a value of 3–5, and when r_k>T, it is determined to be an anomalous feature point; Supplemented by geometric consistency criteria: if the measured spacing of the belt layer steel wires deviates from the CAD value d_cad by more than ±5%, or the RMS residual of the bead ring fitting exceeds the threshold, the entire tire is deemed unusable. The model is estimated by randomly sampling the minimum subset, and the number of interior points (r_k≤T) in all feature points is counted. The model with the most interior points and the corresponding interior point set are retained. This process eliminates abnormal feature points caused by bubbles, defects, noise, overlapping steel wires on the tire shoulder, etc., to avoid contaminating the subsequent fine solution. S353, Levberg–Marquardt nonlinear refined solution: Construct the residual equation system r_k(θ) using only the interior point set retained in step S352, and solve it iteratively with θ0 as the initial value, so that the global cost function Minimum; Solving the incremental equation at each step Where J is the Jacobian of the residuals with respect to θ, obtained analytically or through finite differences; λ adapts to the residuals, increasing λ for larger residuals and biasing towards gradient descent to ensure stability, and decreasing λ for smaller residuals and biasing towards Gauss-Newton to accelerate convergence; the parameter covariance can be derived from (J... J) - ¹Estimate its uncertainty; Convergence criterion: When Stop when the number of iterations reaches the upper limit (e.g., 1e-6) or when the number of iterations reaches the upper limit (e.g., 50 times), and obtain the optimal parameter θ. ; S354, Output: From θ The physical equivalent of each pixel in the five regions is calculated as s_i = p_d·SOD_i / SDD, with the unit being mm / pixel. A pixel-wise distortion correction matrix is generated from the distortion coefficients k1, k2, k3, p1, and p2. The entire solution process is performed automatically without manual intervention. S36, Update the closed loop Write the five-region pixel physical equivalent s_i and distortion correction matrix obtained in step S35 into the device database calibration lookup table to replace the old values; subsequent online detection can directly call them without stopping the machine.
[0038] To verify whether the updated parameters have stabilized, the current (or several) tires are retested using the updated parameters. The deviation between the online measured physical dimensions and the true values of the CAD design is calculated (e.g., root mean square error RMSE or maximum absolute deviation). If this deviation is less than the preset tolerance T (e.g., 0.05 mm), or the relative rate of change of the deviation between two adjacent tires... If the deviation is 1%, the self-calibration is completed and the parameter is locked. Otherwise, the result of this round of retesting is used as the new truth constraint, and the process returns to steps S4 and S5 to enter the next round of iteration until the convergence condition is met or the maximum number of iterations K_max (e.g., 5 times) is reached, at which point the parameter is forcibly locked.
[0039] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, evolutions, or improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope of the present invention.
Claims
1. A method for self-calibration of intrinsic characteristics of tire X-ray images, characterized in that, include: Offline global pre-calibration steps: Obtain the pixel physical equivalents and distortion matrices corresponding to the five regions of the tire as initial calibration parameters; Online real-time size calibration steps: After distortion correction, physical equivalent conversion of pixels in each region, and dual-projection symmetry correction, the detection results are output; Online self-updating calibration steps: Trigger the calibration process, construct and solve the global distortion fitting equation based on the inherent reference characteristics of the tire and the theoretical dimensions of the tire design, obtain the updated calibration parameters and update them online, and achieve parameter locking or iterative optimization through convergence judgment.
2. The tire x-ray image intrinsic feature self-calibration method according to claim 1, characterized in that, In the offline global pre-calibration step, the five tire regions are divided according to the radial structure of the tire by geometric or semantic segmentation methods. The five tire regions include the left sidewall, left shoulder, crown, right shoulder, and right sidewall in sequence along the tread.
3. The tire x-ray image intrinsic feature self-calibration method according to claim 1, characterized in that, In the online real-time dimensional calibration process, the dual-projection symmetry correction step includes: acquiring X-ray images of the tire in two rotational poses of 0° and 180°, measuring the pixel equivalent of each inherent reference feature of the tire in the 0° and 180° poses respectively, and calculating the symmetry mean.
4. The tire x-ray image intrinsic feature self-calibration method according to claim 1, characterized in that, In the online self-updating calibration process, the triggering condition for triggering the calibration process is at least one of the following: timed triggering, tire specification change triggering, or multiple consecutive sets of dimensional deviations exceeding the threshold triggering; the inherent reference characteristics of the tire include the bead wire bundle, the parallel steel wires of the belt layer, and the sidewall and carcass cords.
5. The tire x-ray image intrinsic feature self-calibration method according to claim 1, characterized in that, The inherent reference characteristics of a tire are obtained using the following methods: Denoising and grayscale normalization preprocessing were performed on tire X-ray images; Each inherent reference feature is extracted separately, and the inherent reference features of the tire are separated from the background; Extract the centerline of the inherent reference feature and perform sub-pixel level fine localization of the edge of the inherent reference feature; Geometric fitting yields measured data of the tire's inherent reference characteristics.
6. The tire x-ray image intrinsic feature self-calibration method according to claim 1, characterized in that, Constructing the global distortion fitting equation includes the following steps: The distortion model, cone-beam magnification model, and calibration constraints are combined and incorporated into a projection model described by parameter vector θ, which is defined as follows: , In the formula, Introduced only as a symbol of unknown quantity; SDD represents the distance from the ray source to the detector, which is a constant inherent to the device; p_d represents the inherent pixel pitch of the detector, with the unit of mm / pixel, which is a constant inherent to the device; SOD1 represents the distance from the ray source to the left sidewall; SOD2 represents the distance from the ray source to the left shoulder; SOD3 represents the distance from the ray source to the crown; SOD4 represents the distance from the ray source to the right shoulder; SOD5 represents the distance from the ray source to the right sidewall; x represents the lateral offset of the pixel coordinate; y represents the longitudinal offset of the pixel coordinate; k1, k2, k3 are radial distortion coefficients; p1, p2 are tangential distortion coefficients; The design theoretical dimension coordinates of the inherent reference features of the tire are projected forward through the projection model to obtain the predicted image points, which are then compared with the measured image points to construct the residual vector and the global cost function. When the number of features is greater than the number of unknowns, an overdetermined fitting equation is formed.
7. The method of self-calibration of tire X-ray image intrinsic characteristics according to claim 6, characterized in that, The distortion model is used to correct detector edge distortion. The Brown-Conrady model is adopted, and the formula is: , , In the formula, x_d represents the normalized abscissa of the distorted image plane; x represents the normalized abscissa of the image plane; r represents the radial distance from the normalized image plane coordinate to the principal point; and y represents the normalized ordinate of the image plane. y_d represents the normalized ordinate of the image plane after distortion.
8. The tire x-ray image intrinsic feature self-calibration method according to claim 6, characterized in that, The cone-beam amplification model is M_i = SDD / SOD_i, s_i=p_d / M_i=p_d·SOD_i / SDD, In the formula, M_i represents the geometric magnification; SOD_i represents the distance from the ray source to the plane of the object in the i-th region, i = left tire sidewall, left tire shoulder, tire crown, right tire shoulder, right tire sidewall; s_i represents the model value of the physical equivalent of the pixel in the i-th region, in mm / pixel; p_d represents the detector's inherent pixel pitch, in mm / pixel, which is an inherent constant of the device.
9. The tire x-ray image intrinsic feature self-calibration method according to claim 6, characterized in that, The calibration constraint formula is: s_i'=L_{cad,k} / N_k, In the formula, s_i' represents the measured value of the physical equivalent of the pixel in the i-th region, in mm / pixel; L_{cad,k} represents the theoretical design size of the k-th inherent reference feature in the i-th region; and N_k represents the measured number of pixels on the image.
10. The tire x-ray image intrinsic feature self-calibration method of claim 1, wherein, Convergence criteria include: The tire was retested using the updated calibration parameters, and the deviation between the measured values and the theoretical tire design dimensions was calculated. If the deviation is less than the preset tolerance or the relative rate of change of the deviations of two adjacent rounds is less than the preset threshold, then the deviation is determined to have converged, self-calibration is completed, and the parameters are locked. Otherwise, the retest results will be used as new calibration constraints to enter the next iteration until the convergence condition is met or the maximum number of iterations is reached, at which point the system will be forcibly locked.
