52results about How to "Improve registration accuracy" patented technology

The invention provides a three-dimensional organ reconstruction registration fusion method based on a multi-modal medical image, and belongs to the technical field of medical images. Comprising the following steps: S1, collecting a multi-modal medical image of an ith organ of a target patient in real time, and constructing a preoperative data set; s2, after the first data set is constructed, continuously constructing an intra-operative data set through the intra-operative ultrasonic image; s3, constructing an AI model by using a convolutional neural network; s4, based on the preoperative data set and the intraoperative data set, constructing a stability coefficient and a crowding degree coefficient of the ith organ; and S5, associating the stability coefficient of the ith organ with the crowding degree coefficient of the ith organ, and constructing a comprehensive dynamic risk coefficient. According to the method, CT, MRI and intraoperative ultrasonic images are collected at the same time, the preoperative data set is constructed, the organ volume change rate and the body position adjustment displacement are updated in real time in the operation, the optimal operation path can be automatically predicted, and the subjectivity of depending on doctor experience in traditional operation path planning is reduced.

The invention belongs to the related technical field of aero-engine manufacturing and precision repair, and discloses a blade repair path follow-up compensation method and device based on non-rigid matching, and the method comprises the steps: S1, reconstructing a machining error reference point cloud curved surface based on a scanning data point set and a repair region design model; s2, performing unified modeling on the static deviation of the machining error of the to-be-compensated blade to obtain an error transfer model, and simplifying the corresponding nonlinear comprehensive machining error into a calculation equation of linear pose deviation of a local workpiece coordinate system; s3, establishing a tangent contact approximation cost function based on non-rigid matching so as to register and approach a target curved surface under the constraint of retaining tool path features, and meanwhile, solving a calculation equation by adopting a least square method so as to obtain a linear pose deviation of a local workpiece coordinate system, solving to obtain a pose deviation of each tangent contact coordinate system so as to obtain a linear pose deviation of each tangent contact coordinate system; and furthermore, each cutting contact and each cutter location point are updated, and the shape follow-up compensation of the blade is realized. The compensation precision is improved.

This invention relates to the field of image registration technology, specifically to a dynamic monitoring system for pulmonary nodules using dose-dependent CT scans. The invention obtains an edge cluster encompassing all lung images based on the positional distances between corresponding contour points on edge lines in all lung images; it obtains multiple sets of isotopic contour points based on the positional distribution of all contour points in different lung images; it obtains the referenceability of each set of isotopic contour points by considering the positional relationship between each set of isotopic contour points and other sets of isotopic contour points with the same number of angles on the lung images, as well as the fluctuation characteristics of the corresponding angles of each set of isotopic contour points, and selects anchor points; it constructs a relative positional model of the nodule region in each lung image based on the positional distribution of anchor points and nodule regions in each lung image; and it achieves lung image registration for dynamic monitoring of pulmonary nodules. This invention improves the accuracy of dynamic monitoring of pulmonary nodules by flexibly registering the patient's lung images.

The application discloses a CT-guided old brain PET image spatial standardization system, method and device, the system comprises a PET / CT image acquisition and format conversion module, a PET / CT preprocessing module, a PET / CT image coarse standardization module, a PET / CT image fine standardization module and an automatic SUV extraction module, the old brain PET molecular image spatial standardization is carried out by using a low-dose CT brain structure image as an auxiliary, and the SUV of a region of interest is automatically extracted by using a standard space brain atlas partitioning, the application adopts a two-step strategy from coarse to fine, and by means of an optimized process, an accurate, stable and user-friendly spatial standardization system is provided for the old brain PET image, and the application has important clinical application value and scientific research significance.

This invention discloses an adaptive template registration method based on image grayscale. First, FAST corner points are extracted from the image to be registered. Then, the corner points are clustered, and the minimum bounding rectangle of each set is extracted. The region enclosed by the minimum bounding rectangle in the image to be registered is used as the template to complete adaptive template extraction. Next, a normalized cross-correlation algorithm is used for matching in subsequent image sequences. Finally, the affine transformation coefficients of the image are calculated using the coordinates before and after template matching, and the image to be registered is subjected to an affine transformation to achieve image registration. This invention, through its grayscale-based adaptive template image registration method, provides an adaptive template selection method under different background conditions, solving the problems of fixed template position and limited template information in previous methods, improving the template matching accuracy, and showing good application prospects.

The invention relates to the technical field of data integration, in particular to a coal mine hidden fault recognition-oriented multi-source geological data fusion analysis system, which comprises a response delay section construction module, a seismic response time sequence wave crest interval construction module, a seismic response time sequence wave crest interval construction module and a seismic response time sequence wave crest interval construction module, the direction structure balance regulation and control module adjusts vector density according to a drift trajectory of a borehole diameter symmetry axis to generate balanced direction vector distribution, and the coupling conflict blocking control module reduces conflict node weights based on multi-physical-quantity synchronous inversion characteristics to construct a non-interference data access path. According to the method, timing sequence coherence is determined by dividing a torque fitting interval, vector density is adjusted according to a drift trajectory to maintain space structure balance, conflict node weights are reduced based on synchronous inversion characteristics to block abnormal diffusion, and synchronous feedforward control of timing sequence, direction and structure is realized in an access stage. And the physical authenticity and the registration precision of multi-source geological data fusion are obviously improved.

The application discloses a railway turnout curve-based inspection abnormality early warning method and system, relates to the technical field of track detection and early warning, and comprises the following steps: acquiring three-dimensional point cloud data of a railway turnout region track, performing voxelization processing and outlier filtering on the three-dimensional point cloud data, and obtaining a screening set; based on the screening set, generating a neighborhood point set through a preset neighborhood radius, performing convolution operation, outputting a reference track point cloud feature matrix, acquiring a current track point cloud feature matrix, taking the reference and current track point cloud feature matrices as inputs, and performing a parallel screening mechanism to determine a final reference point; based on the final reference point, solving track geometric offset error, performing clustering, generating a track center line using a cubic B-spline curve, performing turnout curve curvature analysis, and constructing a turnout curve parameter vector; and the application can realize real-time monitoring of track change trends and efficient and accurate inspection of railway turnout curves.

The application relates to the technical field of optical imaging, and discloses an optical lens splicing imaging method and system, the system comprising a calibration module, a preprocessing module, a registration module, a correction module and a fusion module; the registration module, the correction module and the fusion module are arranged, a mixed strategy of global mutual information guidance and local robust matching is used, registration error problems caused by sparse feature points in low-texture areas are solved, the registration precision is effectively improved, a global model based on a Zernike polynomial is used to compensate optical aberration of each sub-lens, the image quality of an edge area of a spliced image is effectively improved, a multi-resolution pyramid optimization mode is used to reduce the calculation complexity of global registration, real-time imaging requirements are met, and a splicing seam is effectively eliminated, and the overall smoothness of an image is significantly improved.

The invention provides a matching method and system for a building white pattern and a time sequence InSAR deformation point, and belongs to the technical field of urban building health monitoring. The method comprises the following steps: obtaining building white mold data of a target area and PS point data of PS-InSAR deformation monitoring; performing vertical registration on PS points based on the PS point data and the elevation information of the building white mold data; performing horizontal registration on PS points based on the plane distribution condition of the PS point data and the building white pattern data; taking the high-rise building in the target area as a research object, judging the visibility of the facade of each high-rise building according to the included angle between the facade of the building and the sight direction of the satellite, and determining the visible facade of each high-rise building; and moving the PS point after vertical registration and horizontal registration to the visible facade of the corresponding high-rise building, and outputting a matching result. According to the invention, accurate matching of the PS point and the building white mold can be realized.

This invention discloses a method for registering industrial printed images based on a lightweight network. The method includes: S1 Image acquisition and dataset construction, which uses a CCD industrial camera to acquire images of the front and back of printed materials and constructs a training dataset by combining multispectral images; S2 Constructing a DHSiam network, which includes two parallel feature extraction networks; S3 Proposing an attention heterogeneous mechanism, which includes a position regression unit and a lightweight unit to enhance data representation and reduce model parameters; S4 Constructing a cross-entropy loss function with additional margin to enhance network performance metrics. This invention builds a hardware platform with 8 parallel GPUs, uses only AMC-Loss to distinguish image similarity, utilizes the attention heterogeneous mechanism to extract important information from feature maps at different scales, and adjusts model parameters to enhance the algorithm's performance on mobile devices, thereby comprehensively improving the model's ability to register printed images with different perspectives and deformations.

The application discloses a kind of based on wheel three-dimensional point cloud measurement and wheel set size method and system, it is related to wheel dynamic detection technical field, including: obtaining point cloud data, and based on point cloud data is carried out wheel outer side surface fitting processing, to fit wheel outer side surface;Extract wheel tread and the junction of wheel outer side surface corner point set;Corner point set is projected to wheel outer side surface and is carried out projection point conversion, to fit plane circle;Determine rolling axis based on the normal vector of wheel outer side surface and the center coordinates of projection point;The center coordinates of projection point is translated along the normal vector direction of wheel outer side surface, to obtain rolling circle data point set;Wheel section contour line is extracted based on rolling axis and rolling circle data point set, to obtain rim parameter.The application can improve the measurement accuracy, and solve the problem that wheel section contour point is unevenly distributed and exists breakpoint.

The application belongs to the technical field of computer vision, and discloses a three-dimensional point cloud registration method based on point-line feature fusion. The three-dimensional point cloud registration method firstly uses SAC-IA to complete static registration of three-dimensional point clouds on the surface of a target object, and establishes an initial motion transformation relationship from the three-dimensional point clouds of the 3D model of the target object to the three-dimensional point clouds on the surface of the target object. Then, in the dynamic tracking stage, point features extracted by ORB are matched with line segment features extracted by LSD-LBD, coarse registration is realized by combining an EPnPL algorithm, and the robustness of pose estimation is improved by eliminating false matching through RANSAC. Finally, in the fine registration stage, point-to-plane ICP is used to further optimize pose estimation. The method does not need to rely on artificial marking, and can complete stable three-dimensional point cloud registration based on natural features, and can be widely applied to complex scenes such as mechanical product maintenance and industrial assembly.

This invention discloses an integrated sensing, storage, and computing linear array remote sensing system and its data processing method. The integrated sensing, storage, and computing linear array remote sensing system includes a host system, a slave system, a housing, and a standard reflector. The host system includes a central processing unit, a data storage device, a high-precision pose module, and an incident light calibration module. The high-precision pose module is used to acquire real-time geographical location and attitude information of the linear array remote sensing system during data acquisition while the external carrier is moving. The host system is installed on the back of the housing. The front of the housing includes several compartments. The slave system includes a linear array imaging sensor and a coprocessor, which are located in the compartments and connected to the host system via an electrical bus within the housing. The host system is used to control the slave system to acquire and process linear array remote sensing data. The standard reflector is used for the calibration of the linear array imaging sensor. The data processing method is used for real-time calibration and mapping of the linear array remote sensing data.

This disclosure relates to a method and model for determining the homography of multimodal images, belonging to the field of image registration technology. The method involves: embedding an initial noisy image into a diffusion denoising process to generate an initial intermediate image; using a cross-modal reference image as guidance, progressively denoising the initial intermediate image based on the diffusion denoising process to generate intermediate images at multiple intermediate steps; and performing complete denoising on the intermediate image at the current time step to generate a noise-free approximate image at the current time step; predicting the increment matrix at the current time step based on the cross-modal reference image and the noise-free approximate image at the current time step; selecting the increment matrix with the smallest registration error from the increment matrices at multiple time steps as the registration increment matrix, and determining the homography matrix of the time step corresponding to the registration increment matrix as the registration homography matrix. The above method for determining the registration homography of multimodal images is used to determine the registration homography matrix.

This invention discloses a regional neural block path planning method based on three-dimensional medical images, relating to the field of neural block technology. The method includes the following planning steps: S1, determining the body posture of the object to be blocked according to the needs of neural block, and acquiring target area scan data through CT and MRI; S2, training the neural system images using a convolutional neural network and a loss function, performing dynamic optimization, outputting the registration results, and constructing an initial model of nerves, blood vessels, and bones. This invention constructs a relevant model of neural block under three-dimensional medical images. By calculating image loss, registration error, and individual adaptation-related data during the registration process, it helps to evaluate and optimize the registration results, reduce registration errors, construct a variable model, and observe and predict in advance, providing good preliminary preparation for regional neural block path planning, thereby improving the accuracy of regional neural block.

This application relates to a bus route and stop registration method, system, storage medium, and intelligent terminal, belonging to the field of computer analysis of public transport networks. It includes acquiring bus route and road network data from different sources; identifying missing parts of the road network based on the bus route data to obtain complete road network data; using a preset spatial clustering algorithm to cluster and merge bus stops with the same station but different coordinates to obtain real bus stops; drawing perpendiculars from the real bus stops to the roads corresponding to the complete road network data to obtain the actual nodes of the bus stops in the road network; and obtaining the bus route nodes and stops located within the complete road network data based on the bus route data, and outputting them as registration points. This application has the effect of clustering and merging bus stops from different data, and then drawing perpendiculars to the real bus stops, so that the registered bus stops can restore the actual situation as much as possible, ensuring the accuracy of the data.

This application discloses a portable autofocus hyperspectral imaging system and method. The system includes a spectral data acquisition module, a spectral analysis module, a displacement platform, an inertial measurement module, and a control module. The spectral data acquisition module includes an adjustable focus lens, a beam splitter, a first collimating lens group, and an image sensor. The spectral analysis module is equipped with an electronic shutter, a slit, a second collimating lens group, a grating, a third collimating lens group, and an imaging element along the optical path. The displacement platform drives the spectral analysis module to perform internal scanning motion. The control module executes the autofocus algorithm, automatic exposure control, and multi-source data fusion and preprocessing. This application can improve focusing accuracy and spatial registration accuracy of spectral data in complex depth-of-field scenes, intelligently and adaptively adjust exposure parameters, compensate for geometric distortion, support unattended continuous operation, reduce the cost of ineffective operations, and improve the reliability and usability of hyperspectral data.

The invention relates to the technical field of hyperspectral remote sensing image processing, and particularly provides a spectral image registration method based on benchmark transmission, which comprises the following steps of: preprocessing a spectral image sequence, selecting a reference image according to gradient information entropy, extracting and matching feature points of spectral images of adjacent spectral bands, and obtaining a spectral image sequence; calculating a homography transformation matrix of the spectral images of the adjacent spectral segments; a direct registration path and a reference transfer registration path are designed, an actual registration path is preferentially selected according to a registration error, and registration of spectral images of all spectral bands is completed through iteration. According to the invention, a special registration process method is designed for the aviation hyperspectral image, the registration efficiency and accuracy are improved, and the problem that the feature point extraction and descriptor of the spectral image are inconsistent in the traditional registration process is solved.

This invention discloses a hand-eye calibration method based on the registration of a planar polygonal contour point set. The steps are as follows: S1, design a calibration reference object with a polygonal contour; S2, use a line laser contour sensor on the end flange of a robot to perform multi-point scanning on each edge contour at different poses to collect multiple point cloud contour data; S3, set the ideal target position of the calibration reference object, construct a target vertex set, and initialize the hand-eye calibration parameters; S4, construct a fitting vertex set of the polygonal contour based on the point cloud contour data, and perform coarse registration with the target vertex set; S5, construct a coarse registration contour feature point set and a coarse registration projection feature point set based on the ideal target position, and obtain the optimal solution for the hand-eye calibration parameters by minimizing the distance difference between the matching point pairs of the two; S6, design a hand-eye calibration performance evaluation index to evaluate and select the optimal solution for the hand-eye calibration parameters. This method has high registration efficiency and accuracy, low cost, and good stability.

This invention discloses a joint sensor registration and multi-target tracking method based on an event-triggered mechanism. Addressing the problems of large computational redundancy and lack of adaptability caused by periodic updates in existing methods, this invention constructs registration errors based on the posterior distribution of multiple targets and further builds event-triggered statistics. Threshold decision and hysteresis mechanisms enable on-demand triggering updates during the registration process, thereby avoiding unnecessary repetitive calculations. Furthermore, spatial transformation parameters between sensors are updated and coordinate alignment is completed only when trigger conditions are met. Subsequently, multi-sensor information fusion is performed to obtain consistent multi-target state estimation results. To improve registration stability, group-level statistical features can be incorporated into the model. This invention significantly reduces computational complexity through the event-triggered mechanism, improves the stability and robustness of the registration process, and enhances adaptability to dynamic environmental changes.

The invention discloses a Landsat 8 / 9 cross-spectrum sub-pixel registration method and system based on structural similarity constraint, relates to the field of remote sensing image processing and space geometric correction, and aims to solve the problem of unstable geometric alignment caused by radiation difference, unreliable region interference and similarity measurement failure in cross-spectrum sub-pixel registration. The method is technically characterized by comprising the following steps of: 1, converting a to-be-registered Landsat 8 / 9 cross-spectral image into a structure representation space, and constructing a reliability weight map; 2, carrying out frequency domain phase correlation analysis, carrying out bandwidth limitation and window processing on a frequency spectrum to obtain a stable correlation peak, and obtaining a sub-pixel level translation initial estimation value; 3, constructing a weighted structural similarity objective function in the structural representation space, optimizing translation parameters through iteration, and inhibiting the influence of an unreliable region on the objective function by using a reliability weight; and 4, outputting a global or block-level sub-pixel displacement estimation result, and calculating a confidence or uncertainty index.

This invention relates to the field of computer vision technology, specifically to a point cloud registration method, system, storage medium, and apparatus based on a diffusion model. The method includes: acquiring training samples, including source point clouds and target point clouds; using a point cloud feature extraction module to extract features from the source and target point clouds respectively, obtaining features 1 and 2; using a feature interaction module to fuse features 1 and 2, obtaining features F1 and F2; using a point cloud generation module to generate two aligned point clouds, Ag and Bg; using a transformation matrix prediction module to calculate the correspondence between the original point cloud and the generated point cloud Ag, and between the original point cloud and the generated point cloud Bg, obtaining the predicted transformation matrix and translation vector; calculating the model's loss function, adjusting the model parameters, and completing model training when the loss function converges. This invention, based on a diffusion model, transforms the registration problem into a generation problem, improving the accuracy of point cloud registration.

The application discloses a point cloud registration method based on global feature and local feature fusion, comprising the following steps: obtaining a reference point cloud and a point cloud to be registered; extracting point cloud features of the reference point cloud and the point cloud to be registered respectively to form an initial reference point cloud feature set and an initial point cloud to be registered feature set; introducing self-attention mechanism and cross-attention mechanism, and introducing correlation information between the reference point cloud and the point cloud to be registered to form a final reference point cloud feature set and a final point cloud to be registered feature set; obtaining similarity information between each point pair; constructing a matching point pair relationship matrix according to the similarity information between the point pairs; obtaining a transformation model parameter of the reference point cloud and the point cloud to be registered; and registering the reference point cloud and the point cloud to be registered. The application fully excavates potential spatial topological structure information of point cloud data by combining local features and global features, obtains point cloud fusion features with rich semantic information, and can effectively improve the accuracy of point cloud registration.

The application discloses a vehicle-infrastructure collaborative positioning method based on multi-view degradation perception, relates to the technical field of intelligent transportation and automatic driving, and comprises the following steps: acquiring vehicle end point cloud data, and constructing a vehicle end point-surface residual error factor; acquiring measurement data of a vehicle-mounted inertial measurement unit, and constructing an IMU pre-integration factor through mechanical arrangement; acquiring roadside point cloud data, constructing a roadside residual error based on the roadside point cloud data and the vehicle end point cloud data, and constructing a roadside DICP factor; constructing a factor graph model based on the vehicle end point-surface residual error factor, the IMU pre-integration factor and the roadside DICP factor, updating the pose information of a vehicle, and obtaining a positioning result. The application can effectively overcome the defects of drift and unstable registration caused by the insufficient observability of a laser radar in a degradation scene, solve the problems of high delay and insufficient constraints of multi-view vehicle-road collaborative positioning, and realize long-time, stable and high-precision real-time positioning in a complex traffic environment.

This application discloses an image generation method, apparatus, device, medium, and program product. By physically contacting the probe of an acquisition tool with the spinous process feature point of the first vertebra of a skeleton, information on the relative positional relationship between the acquisition tool and the spinous process positioning frame is obtained. Based on the relative positional relationship information, the position information of the acquisition tool in the coordinate system of the spinous process positioning frame is determined. The position information of the first vertebra of the skeleton in the preset vertebral coordinate system is transferred to the coordinate system of the spinous process positioning frame to obtain the position information of the first vertebra in the coordinate system of the spinous process positioning frame. The position information of the spinous process positioning frame and the target plate is acquired in real time, and the posture of the target plate is adjusted so that the central area of ​​the target plate in the coordinate system of the spinous process positioning frame is aligned with the center position of the first vertebra. The posture of the camera is adjusted so that the camera focus is aligned with the central area of ​​the target plate. The camera is used to take a picture of the skeleton to obtain a target image. According to the embodiments of this application, image quality and registration accuracy can be improved.

The application discloses a target edge distance measurement method based on fusion of vision and laser radar, relates to the technical field of industrial vision detection, and comprises the following steps: laser radar and a camera are jointly calibrated through a calibration point cloud and a calibration board image, and a transformation matrix of the laser radar and the camera is obtained through feature matching and spatial relationship solving; based on the transformation matrix, original point cloud data is converted from a laser radar coordinate system to a camera coordinate system, then is mapped to an image pixel plane, and point cloud projection pixel coordinates are generated; according to a target edge ROI region, the point cloud projection pixel coordinates are screened, and ROI associated point cloud data is generated; the ROI associated point cloud data is preprocessed, and pure point cloud data is generated; the geometric center distance and the robust quantile distance of the pure point cloud data are calculated, and single-frame distance data is generated; the single-frame distance data of continuous multiple frames is integrated, the distance between a target to be measured is calculated, threshold and stable frame number verification are then performed, and the measured distance of the target to be measured is output.

The invention relates to a multispectral image registration method, in particular to a multispectral image registration method based on improved SIFT (Scale Invariant Feature Transform) features. The method comprises the following steps: step 1, preprocessing multispectral remote sensing image data, wherein the multispectral remote sensing image data comprises a reference multispectral image and a to-be-registered multispectral image; step 2, extracting image gradient features based on the preprocessed multispectral remote sensing image data; extracting edge corner point features based on image gradient features; step 3, extracting an improved SIFT feature based on the image gradient feature and the edge corner feature; 4, performing homonymy point matching calculation on the reference multispectral image and the multispectral image to be registered based on the improved SIFT features to obtain a transformation matrix; and step 5, converting the multispectral image to be registered into the reference multispectral image based on the transformation matrix to obtain a final registration result. According to the method, the number of effective feature points in the image registration process can be increased, and rapid and accurate image registration is realized.

The invention discloses an insulator aging intelligent detection method and system based on the hyperspectrum of an unmanned aerial vehicle. The method comprises the following steps: completing synchronous calibration of radiation, spectrum, camera internal reference and time; flying according to a preset route, synchronously collecting data, and unifying timestamp marking; performing radiation correction on the hyperspectral data, and performing distortion correction on the visible light image; establishing a one-to-one correspondence relationship of the data through timestamp matching; based on the visible light image feature points, an insulator string sparse three-dimensional point cloud is constructed, and coarse registration of hyperspectral data and the three-dimensional point cloud is achieved; iteratively optimizing by taking the insulator feature point as a reference, correcting topographic deviation according to a digital elevation model, and ensuring that a registration error meets a threshold pixel; segmenting and extracting a single insulator, establishing a local coordinate system, performing geometric correction according to an actual space attitude, and eliminating inclination deflection distortion; and extracting the spectral characteristics and three-dimensional coordinates of the corrected insulator, judging the aging grade, and outputting a detection report containing accurate positioning.

This invention relates to a non-rigid registration method and system based on the bending morphology of a puncture needle. The method includes acquiring the three-dimensional bending morphology of the puncture needle within the tissue in real time using multimodal sensing technology to generate needle morphology data; performing initial rigid registration based on preoperative CT and intraoperative space based on the data; performing non-rigid fine registration on this basis to compensate for errors caused by tissue deformation and physiological movement; fusing multi-source sensor data for state estimation and error compensation to generate dynamic registration results; and using the results to drive a robot to achieve closed-loop control of the puncture path. The system includes multimodal sensors, a registration calculation unit, a multi-source data fusion unit, and a control unit, which collaboratively complete morphology reconstruction, registration calculation, and dynamic guidance. This invention can significantly improve the registration accuracy and stability of puncture surgery, reduce the risks caused by target displacement, and is applicable to various robot-assisted puncture scenarios.

This invention provides a method, apparatus, and electronic device for image registration of rail vehicles. The method includes: inputting an image to be detected into a first detection model; detecting Regions of Interest (ROIs) including key components in the image using the first detection model to obtain a first coordinate set corresponding to the ROIs; obtaining a first transformation matrix based on the first coordinate set and a first reference coordinate set; performing a first registration based on the first transformation matrix; inputting the first-registered image into a second detection model; detecting key components in the image using the first detection model to obtain a second coordinate set; obtaining a second transformation matrix based on the second coordinate set and a second reference coordinate set; obtaining an offset reference vector based on the second transformation matrix; determining offset vectors corresponding to at least one key component based on the offset reference vector and a preset tolerance; and performing a second registration on the first-registered image based on the offset vectors. This invention improves image registration performance and increases detection accuracy.