84 results about "Geometric alignment" patented technology

The invention discloses a monocular video face reconstruction and driving method based on Gaussian splashing and a 3DMM model. The monocular video face reconstruction and driving method comprises the following steps: S1, converting each frame of a face RGB video into an explicit expression face grid based on the 3DMM model; s2, constructing a geometric enhancement module, and carrying out geometric alignment on eye and tooth regions by adopting an improved iterative nearest point algorithm; s3, initializing a 3D Gaussian ball, and driving the dynamic change of the 3D Gaussian ball according to the point change of the face grid under different expressions; s4, based on the characteristics of different face areas, adaptive optimization learning strategy coefficients are distributed; and S5, increasing confidence information for each 3D Gaussian ball, generating a confidence image through rendering, multiplying the confidence image by a rendered image, calculating high-dimensional feature similarity with the input image, and optimizing attribute parameters and rendering effects of the 3D Gaussian balls. According to the method, Gaussian splashing and a 3DMM model are combined, and high-precision reconstruction and driving of a monocular video face are realized.

The invention provides a camera external parameter calibration method and device based on image and point cloud matching, and the technical scheme of the invention is that a point cloud rendering view corresponding to an original point cloud is rotated, so that the point cloud rendering view and a camera image are superposed visually, and initial space association is established; and directly learning and solving accurate camera external parameters from the local relevance between the original point cloud intensity information and the image RGB information in an end-to-end manner by using a deep learning model. According to the scheme, the strict camera-radar orientation consistency requirement in a traditional method is not needed through rough matching, manual participation in the whole process of traditional calibration is replaced, transition from visual alignment to geometric alignment is achieved, and the calibration efficiency and scene adaptability are remarkably improved. Meanwhile, the scheme of the invention initiates a process of visual field cone cutting-perspective projection rasterization-intensity-RGB end-to-end matching, and compared with a traditional feature point method, the calculation complexity is greatly reduced, and the memory occupation is reduced by 80%.

The invention provides an intelligent visual inspection method and system for hardware stamping die production, and relates to the field of hardware stamping die production, and the method comprises the steps: obtaining a die current state image of a to-be-detected hardware stamping die and a defect-free die state reference image, and carrying out the geometric alignment; and then calculating a difference graph to obtain significance distribution of a potential defect area, carrying out multi-scale difference feature hierarchical coding by using a pyramid network, and separating normal wear artifacts from real defects. Furthermore, a feature receptive field anchoring mechanism is introduced, and the attention weight is dynamically adjusted according to the geometric topology of the mold. And finally, a lightweight classifier is used to make a decision on the enhanced difference features, and sub-millimeter-level anomaly recognition is realized. In this way, the sensitivity to tiny changes of the hardware stamping die can be effectively improved, and defects can be found in the early stage.

The invention discloses a digital twinning-based virtual travel platform construction method, which relates to the technical field of digital twinning and cultural travel, and comprises the following steps of: initializing an instruction set based on federated learning, dynamically distributing a federated model precision weight through a federated learning framework, and calibrating a real-time video stream and a three-dimensional model by using a video twinning engine to obtain a real-time video stream; generating a virtual-real scene bidirectional mapping protocol containing a precision grading parameter and an NPC control instruction; and based on a virtual-real scene bidirectional mapping protocol, scanning a physical space three-dimensional geometric structure through ray detection, executing a mixed reality collision monitoring algorithm, and generating a closed-loop interaction data set containing an NPC behavior instruction and a physical equipment control signal. According to the method, the precision weight of the federal model is dynamically allocated, a target detection algorithm is triggered by combining the real-time video stream frame rate fluctuation to perform semantic segmentation on a key region, and the geometric alignment error value is calculated to divide the precision level, so that the virtual environment and the real world are seamlessly jointed.

The invention provides a high-fidelity virtual fitting method and system based on multi-modal representation fusion, and belongs to the field of computer image generation, and the method comprises the steps: S1, inputting a figure image and a garment image into an appearance keeping deformation alignment module, and outputting an appearance flow graph # imgabs0 #, thereby achieving the geometric alignment between a garment and a figure, and keeping the appearance details of the garment; s2, constructing a semantic representation and understanding module for providing semantic representation for clothing attributes, generating a corresponding text for a clothing image, and ensuring semantic alignment consistency between clothing and a human body under various human body postures; and S3, constructing a multi-modal priori guided appearance generation module, and integrating the multi-modal features and priori knowledge of the pre-training model to optimize an appearance generation effect and realize consistency and unification of semantics and geometry. According to the method, the clothes can be accurately aligned to the human body, and meanwhile fine-grained detail information of the clothes is reserved.

The invention discloses an automatic texture color mapping method for large-scale point cloud data. The automatic texture color mapping method comprises the following steps: firstly, realizing geometric alignment of geographic texture data and point cloud data; then realizing release of intra-core buffer by means of a mass point cloud space segmentation and retrieval management technology, and segmenting the mass point cloud into a plurality of regional point cloud blocks; carrying out filtering processing on the regional point cloud blocks to filter out discrete flying spots; secondly, considering correction constraints of different texture source information to realize seamless generation of geographic texture data; secondly, efficient point cloud preliminary color assignment is achieved by means of the spatial relation between geometric mapping and local aggregation of regional point cloud blocks and geographic texture data; secondly, eliminating the texture information (color) of the wrong coloring point cloud by using the self-shielding judgment result of the point cloud; and finally, realizing spatial merging of regional point cloud block data sets by using an efficient out-of-core file merging technology. According to the method, the dependence on hardware equipment (such as a large memory and a high graphics card) is avoided, and the block processing of massive point clouds is efficiently realized.

A robot grabbing method based on a geometric topology constraint field and adaptive calibration comprises the following steps: constructing a manifold neighborhood enhancement field based on a target object prior point cloud; carrying out feature coding on an input real-time observation point cloud and a prior point cloud template, and embedding observation features into a manifold neighborhood enhancement field to realize geometric alignment; aiming at the RGB image and the real-time observation point cloud, establishing a preprocessing and symmetry sensing mechanism of a multi-modal feature; constructing a geometric feature re-calibration module, performing adaptive weighting on feature response through a channel attention and space attention mechanism, and performing adaptive calibration on deviation between observation features and prior manifold features; the 6D pose parameters of the target object are regressed based on the re-calibrated global features, and a corresponding geometric consistency measurement value is output; and according to the consistency measurement result, the pose result is subjected to iterative correction in the reasoning stage, and a robot grabbing instruction is output. According to the method, the space postures of different objects can be accurately modeled, and the grabbing precision of the robot in a complex environment is improved.

The invention relates to the technical field of multi-camera image processing and computational photography, and discloses an image processing method and system based on four-camera cross-focal-length continuous zooming fusion. The image processing method comprises the steps of system initialization, image acquisition, zoom routing, automatic ROI extraction and tracking, field-of-view cutting and geometric alignment, image fusion and binocular depth recognition. Through systematized multi-camera collaborative design, an innovative mechanism is introduced in key links such as zoom routing, geometric alignment, image fusion and depth recognition, smooth zoom, space consistency, detail fidelity, power consumption optimization and high-quality 3D perception are realized, the imaging quality is improved through the effects, the application scene is expanded, and the application prospect is wide. And a comprehensive solution is provided for mobile photography, AR and intelligent visual systems.

The invention relates to a multi-modal data fusion method, system and device and a medium. The method comprises the following steps: acquiring synchronous RGB (Red, Green and Blue) images and laser radar point cloud data; geometric alignment is carried out on the point cloud data, and a mapping relation between the point cloud data and image pixels is established; respectively extracting two-dimensional visual features of the image and three-dimensional geometric features of the point cloud based on the mapping relation, and re-projecting the three-dimensional features to a two-dimensional space aligned with the visual features; and finally, dynamic weighted fusion is carried out on the two types of features through an adaptive attention mechanism, and a multi-modal fusion feature map is generated. By adopting the method, the fusion weight can be automatically adjusted according to the environment change, and the robustness and accuracy of a sensing system in a complex scene are effectively improved.

The invention discloses an indoor multi-mode 3D target detection method based on Transform, and belongs to the technical field of artificial intelligence and computer vision. According to the method, point cloud data and RGB image data are fused, and an innovative Geo-ResAlign space alignment module and a FusionNet multi-modal fusion module are combined, so that the problems of low detection precision and poor robustness caused by sparse point cloud data, multi-modal alignment errors and insufficient feature utilization in a traditional method are effectively solved. In specific implementation, firstly, preprocessing and feature extraction are carried out on point cloud and image data; then, through an inverse geometric alignment mechanism and a residual error alignment mechanism of a Geo-ResAlign module, the spatial deviation of the multi-modal data is eliminated; carrying out cross-modal feature fusion by utilizing a Fusion Net module, and realizing deep interaction through confidence coefficient screening, dynamic weighting and a Transform self-attention mechanism; and finally generating a 3D bounding box and target category information. The method remarkably improves the detection precision and efficiency in a complex indoor scene, and can be widely applied to the fields of robot navigation, intelligent security and protection, industrial automation and the like.

The invention provides an engineering project management method and system based on a rule engine. The engineering project management method comprises the following steps: acquiring an image of an information data scanning file or acquiring the image through a CCD (Charge Coupled Device) camera; performing preprocessing, geometric alignment and correction on the image, extracting a global embedded vector by using a ResNet pre-training model, and finishing alignment through perspective transformation in combination with an improved cosine similarity matching template; extracting text content by utilizing an OCR (Optical Character Recognition) technology, matching a field format through a regular expression, and converting a text into structured data; and checking the structured data by using a rule engine, generating a question list, and sending the question list to a user. The method improves the efficiency and accuracy of information data inspection, is suitable for various document scenes such as contracts, bills and receipts, and has high adaptability and wide application prospects.

The invention provides a Riemannian graph word segmentation device for structure knowledge migration, which belongs to the field of graph basic models, and comprises a geometric vocabulary sampling module used for sampling input graph data to obtain geometric vocabularies; the geometric vocabulary encoding module is used for respectively mapping geometric vocabularies into corresponding constant curvature Riemannian spaces for coordinate encoding according to different structural modes of the geometric vocabularies to obtain space coordinates corresponding to the structural modes of the geometric vocabularies, discretization is carried out by using the Riemannian quantization module to obtain quantization marks, and finally, the geometric alignment decoding module is used for decoding the geometric vocabularies. And fusing the quantitative marks in different Riemannian spaces to obtain a unified graph structure representation capable of supporting structure knowledge migration. The problems that in the prior art, due to the fact that a curvature selection mechanism of a local form of a graph structure is lacked, a word segmentation device cannot dynamically adapt to geometric characteristics of the structure, and structural representation aliasing exists, the graph structure coding accuracy is insufficient, and the cross-domain knowledge migration generalization ability is low are solved.

The invention discloses a welding seam welded area segmentation method and system based on a semantic segmentation network and depth geometric constraint, and relates to the technical field of welding seam welded area extraction, and the method comprises the steps: obtaining an RGB image and a depth image of a to-be-welded part; acquiring a rectangular detection frame and a preliminary mask for marking a welded area of the weld joint; constructing an energy model; performing global pixel-level optimization on the preliminary mask by adopting an energy model to determine a mask boundary, and obtaining an optimized mask; calculating to obtain a gradient field, a normal direction and a local curvature of the depth image in a boundary neighborhood, and interpolating the depth change curve along the normal direction to obtain an interpolation curve; and performing depth interpolation processing on the optimized mask by utilizing the interpolation curve to obtain three-dimensional boundary information of the welding seam. According to the method provided by the invention, three aspects of feature expression, geometric alignment and engineering availability are considered at the same time, and a complete boundary inference process from network coarse segmentation to deep geometric fine positioning is realized.

The invention discloses a short-focus projection image correction optimization system based on artificial intelligence, and the system comprises an image collection and structure modeling module which is used for generating a structure perception feature map; the structure guide matching module is used for constructing a stable matching subset and executing matching point screening; the dense matching and geometric correction module is used for outputting a coarse correction image; the structure residual error fine tuning module is used for constructing a structure residual error image and inputting the structure residual error image into the local fine tuning sub-network, guiding the model to carry out fine-grained correction on a structure error region, and generating a final correction image; the structure consistency evaluation module is used for evaluating key point deviation, edge direction similarity and geometric alignment degree; the correction convergence judgment module is used for tracking the correction score change; and the correction result output module is used for outputting a final correction image. The method has the advantages of high precision, strong robustness and iterative optimization, and is suitable for image geometric alignment and enhancement requirements in complex optical projection scenes such as smart classrooms, interactive display, immersive projection and the like.

The invention discloses a positioning method for generating a pseudo LiDAR descriptor for cross-modal matching based on visual depth estimation, and the method comprises the following stages: in an offline training stage, obtaining synchronous multi-modal data, inputting a camera image into a pre-trained monocular depth estimation network, generating a dense depth map, and carrying out the alignment; training a global descriptor generation network with weight sharing; a laser LiDAR global descriptor database is constructed; an on-line positioning stage: acquiring a surrounding environment image; inputting the image into a pre-trained monocular depth estimation network to generate dense depth map alignment; inputting a global descriptor generation network, and generating a visual global descriptor; and performing cross-modal descriptor retrieval and matching. Generating a geometrically aligned pseudo LiDAR depth map through monocular depth estimation, and realizing cross-modal matching with a real LiDAR map; the system hardware threshold is obviously reduced; introducing a triple supervision mechanism based on geometric overlapping degree and a yaw angle invariant descriptor network of weight sharing; and a modal gap between the vision and the LiDAR is bridged.

The invention relates to a three-dimensional visual synthesis system based on heterogeneous collaboration and screen alignment fusion, which is characterized in that a master control system generates complete viewpoint configuration information of a current scene acquired by an image acquisition device, and the complete viewpoint configuration information is used as a single data source and is simultaneously distributed to a graphic processing unit and a neural rendering coprocessor of the master control system; synchronous starting and parallel execution of the two heterogeneous rendering pipelines are achieved. And in combination with a subsequent screen alignment fusion mechanism, an end-to-end collaboration framework with low delay, low coupling and high coordination is constructed, and the key problem of real-time integration of high-fidelity contents is effectively solved. On the premise of not depending on a traditional deep buffer area or geometric alignment, high-fidelity visual content generated by special hardware acceleration and a dynamic environment of a main control system are efficiently fused, and complex effects such as lightweight asynchronous communication and real-time shadow are supported. Therefore, multiple bottlenecks in the aspects of real-time performance, compatibility and engineering deployment in the prior art are broken through.

The invention discloses an image restoration method, which comprises the following steps: acquiring a target image, a reference image set and a boundary intensity image, and extracting target features, a reference feature list and boundary features; based on the extracted features, performing similarity modeling by using a learnable spiral path to generate a similarity model result; based on a similarity model result, performing geometric alignment and double-domain fusion on the reference feature list to generate initial fusion features and alignment parameters; calculating an error mode pattern based on the initial fusion features; executing closed-loop feedback according to the error mode pattern, and iteratively updating to generate a final fusion feature; and decoding the final fusion feature to obtain a repaired image. Through spiral sampling of boundary perception, non-folding alignment and closed-loop feedback, the structure consistency and detail fidelity of the repaired image under non-rigid deformation are improved.

The invention discloses a dynamic shelter restoration method and system based on continuous streetscape panoramic images, and the method comprises the steps: firstly obtaining a to-be-restored target panoramic image A and a to-be-restored reference panoramic image B, and generating an original pixel-level shelter mask; secondly, extracting matching points among the panoramic images, realizing cross-view geometric alignment of the images, and obtaining a target perspective view C and a reference perspective view D through perspective re-projection; and then inputting the target perspective view C and the reference perspective view D into a three-dimensional reconstruction framework, and performing three-dimensional point cloud reimaging under the camera pose of the target perspective view C by using a depth inspection mechanism. And finally, carrying out image restoration on a three-dimensional point cloud re-imaging result, restoring to a panoramic coordinate system, splicing with an original panoramic image, and outputting a shielding-free panoramic image. According to the method, it is ensured that the repairing result conforms to the authenticity and geometric consistency of the geographic space, and the problems of overlapping conflicts and visual tearing during multi-view projection fusion are effectively solved.

The invention relates to the technical field of camera shooting. The virtual reality panoramic dynamic shooting method and system based on multi-lens cooperation are provided, and the method comprises the steps: carrying out the exposure time sequence synchronous control of a multi-lens array, and generating multi-channel image data of synchronous exposure; performing noise feature separation processing on the synchronously exposed multichannel image data to obtain a high-frequency noise component and a brightness component of each lens; constructing a multi-channel noise correlation model, carrying out combined suppression processing on the noise of the overlapped region, and generating a noise-reduced multi-channel image; performing geometric alignment and dynamic splicing processing on the multi-channel image after noise reduction to generate a panoramic dynamic image; and carrying out real-time rendering processing on the panoramic dynamic image, and outputting a low-noise panoramic picture with an adaptive view angle, so as to solve the problems of image dislocation and dynamic object ghosting caused by noise accumulation and time sequence jitter in an overlapping region under multi-lens collaboration.

The invention relates to the technical field of unmanned aerial vehicle visual precise landing scheme design, in particular to a method and system for improving the visual precise landing effect of an unmanned aerial vehicle and electronic equipment. A plurality of cameras fixed at four corners of a parking apron are used for synchronously acquiring video streams, and two-dimensional coordinates are mapped to a unified three-dimensional coordinate system by using a multi-view geometric alignment algorithm after anti-illumination interference preprocessing. The dynamic trajectory of the real-time position and attitude of the unmanned aerial vehicle is optimized based on Kalman filtering, and trajectory anomaly judgment is realized in combination with an LSTM prediction model and convex hull safety region detection. And triggering a grading early warning mechanism including emergency braking and light warning when detecting that the position crosses the boundary, the acceleration exceeds the limit or the reconstruction error is abnormal. Positioning robustness in a strong light environment is remarkably enhanced, centimeter-level landing precision is achieved, and collision accidents are effectively prevented through a multi-level safety protection system.

The invention relates to the field of three-dimensional modeling and data set construction of ship outfitting parts, and particularly provides a ship outfitting part multi-angle RGB graph and standard coordinate graph data set construction method. The method comprises the steps of performing direction label analysis and normalization processing on a fitting-out part three-dimensional model based on a deep learning model, optimizing a view angle configuration file and generating multi-view angle configuration; adjusting the intensity and distribution of a light source according to the material characteristics of the fitting-out part, and rendering an RGB image and a standard coordinate graph; analyzing geometric consistency and direction matching of rendering results through multi-modal consistency verification, and generating an abnormal data list; the diversity of the data set is expanded by combining a data enhancement module; and finally, evaluating the multi-view coverage rate and consistency of the data set, and feeding back and optimizing a direction label generation model and a view configuration file to form a closed-loop improved mechanism. According to the method, a high-quality multi-angle RGB image and standard coordinate graph data set can be automatically generated, the consistency and geometric alignment degree of images are ensured, a three-dimensional reconstruction task and deep learning model training are effectively supported, and the quality and adaptability of the data set are improved.

This invention discloses a 3D defect detection method, system, and device based on typical geometric alignment and multi-view adaptive rendering, comprising the following steps: S1, acquiring the initial 3D point cloud of the object to be detected, and performing unsupervised feature analysis on the initial 3D point cloud through typical geometric alignment to obtain a standardized aligned point cloud of the object to be detected; S2, converting the standardized aligned point cloud of the object to be detected into a set of geometric feature maps from multiple camera views; S3, inputting the geometric feature maps of the object to be detected from the camera views into a general 2D image anomaly detection network, outputting point-level anomaly scores, and then back-projecting them onto the corresponding surface points of the standardized aligned point cloud to obtain object-level anomaly indicators for 3D defects. This invention does not rely on the geometric structure and observation posture of the object to be detected, and can achieve robust anomaly defect detection on unaligned object 3D point clouds, achieving high-precision point-level anomaly localization.

The invention belongs to the field of three-dimensional scene understanding, and provides a weak supervision three-dimensional target detection method based on multi-signal decoupling. According to the method, a multi-signal decoupling learning framework is constructed, and the position, the size and the orientation of a three-dimensional bounding box are respectively supervised by utilizing two-dimensional target frame constraint, category size priori generated by a language model and point cloud geometric information. The method specifically comprises a centrality-enhanced projection constraint module used for optimizing target positioning, a semantic prior anchoring module used for guiding size regression, a rotation consistency regularization module used for improving orientation discrimination capability, and an adversarial geometric alignment module used for dynamically adjusting a three-dimensional frame boundary. All the modules are trained cooperatively, a high-quality pseudo label can be generated for detector training without three-dimensional labeling, and finally the high-precision weak supervision three-dimensional target detection method is realized.

According to the multi-task processing model learning method and the multi-task processing execution method using a machine learning model learned based on the method provided by the embodiment of the present invention, multi-tasks for performing output based on a plurality of domains can be processed, and the multi-task model learning method and the multi-task processing execution method using the machine learning model learned based on the method can be provided. According to the multi-task processing model learning method, knowledge data scattered in a potential space according to each task are mutually migrated and learned through geometric alignment in the comprehensive potential space, and the multi-task processing execution method of a machine learning model learned based on the method is utilized.

The invention relates to an unsupervised parallax-tolerant terahertz image splicing method and system, and belongs to the technical field of image processing. Comprising the following steps: S1, a geometric deformation correction stage: carrying out feature extraction, matching and deformation estimation on an input terahertz image pair with low resolution, sparse texture and parallax to realize geometric alignment; s2, an image fusion stage: carrying out seamless fusion on the geometrically aligned image pairs, and optimizing a splicing boundary transition effect; aiming at the characteristics of low resolution, sparse texture, remarkable parallax and the like of a terahertz image, through two-stage architecture and multi-module collaborative optimization, the method is superior to an existing method on self-built terahertz data sets and UDIS-D standard data sets, geometric consistency and seamless fusion effects under unsupervised training are achieved, and the method is suitable for being used in the field of non-supervised training. And the application range of terahertz imaging in large scenes such as industrial detection and safety inspection is effectively expanded.

The application relates to the technical field of image processing, and discloses a multispectral fusion imaging camera module and an image processing method, which comprises the following steps: performing spatial mapping on infrared imaging images and near-infrared imaging images based on the spatial arrangement order of corner points and the direction of continuous edge broken lines; determining a spectral corresponding structure according to the light and dark trend changes of the same scene in three imaging channels; performing regional division on a geometric alignment result based on the spectral corresponding structure; performing direction consistency correction on geometric edges in a visible light imaging image based on a dominant contour line; performing extensibility superposition on texture lines in the spatially mapped near-infrared imaging image based on the texture direction of a texture detail area; performing superposition on the brightness levels of the visible light imaging image based on the brightness extension direction of a background smooth area; and performing connectivity checking on the boundary direction, the texture direction and the brightness extension direction. The application ensures the reliability of multispectral image fusion.

This invention discloses a scale-redistribution-based top-view fisheye pedestrian detection method, belonging to the field of target detection technology. In the feature generation stage, this invention employs frequency-preserving downsampling, concatenating low-frequency contours and high-frequency textures after sub-pixel rearrangement and wavelet decomposition in the channel to achieve spatial depth swapping. In the feature fusion stage, it increases the participation ratio of high-resolution features through adaptive upsampling for geometric alignment. In the spatial computation stage, it divides features into high-response and low-response regions by learning a spatial weight map, performing focused convolution only on high-response regions. In the optimization stage, it constructs a joint regression mechanism, jointly modeling overlap consistency and distribution distance, adjusting sample gradient contributions based on detection difficulty, and controlling the angle supervision intensity according to the target aspect ratio, forming a scale redistribution that spans multiple stages. This invention achieves stable detection of small-scale and distorted targets in top-view fisheye images, improving the detection consistency between edge and dense regions.

This invention relates to the field of automated robot assembly technology, specifically a method and system for robot assembly tasks based on 3D vision. The method includes the following steps: Step 1: Workpiece surface scanning: Scan the workpiece surface, construct a 3D model of the workpiece surface, determine each measurement point n, and construct a point cloud based on the depth measurement values ​​of the measurement points; For the first time, the ICP algorithm is used to solve the robot assembly task by employing a low-cost sensor, that is, the cost function is iteratively minimized based on the estimation of the point correspondence between two point clouds, making the process robust to incorrect geometric alignment. In the subsequent search phase, the Lissajous function is used for surface sliding and admittance control, enabling the robot to complete the assembly task under the condition of random position errors.

This invention discloses a sparse-view 3D reconstruction method based on block-based geometric alignment and texture decoupling, belonging to the field of computer vision and 3D reconstruction technology. The method includes: acquiring RGB images and prior depth maps from a sparse viewpoint, and initializing a 3D Gaussian sputtering model; eliminating monocular depth local scale drift through block-based local alignment and constructing a second-order normal loss to constrain geometric consistency; achieving adaptive smoothing regularization based on texture-structure decoupling weights guided by normal gradients; generating artifact masks through depth residual analysis, applying soft penalties and performing hard pruning on anomalous Gaussian primitives; jointly optimizing color reprojection loss and geometric, smoothing, and penalty losses, iteratively updating Gaussian parameters, and achieving high-fidelity rendering of the new viewpoint. This invention employs the aforementioned sparse-view 3D reconstruction method based on block-based geometric alignment and texture decoupling, solving the problems of geometric instability, texture-structure coupling interference, and floating artifacts under sparse viewpoints.

This invention discloses an image restoration method, comprising: acquiring a target image, a set of reference images, and a boundary intensity map; extracting target features, a list of reference features, and boundary features; based on the extracted features, performing similarity modeling using a learnable spiral path to generate a similarity model result; based on the similarity model result, performing geometric alignment and dual-domain fusion on the list of reference features to generate initial fusion features and alignment parameters; calculating an error pattern map based on the initial fusion features; performing closed-loop feedback according to the error pattern map to iteratively update and generate final fusion features; and decoding the final fusion features to obtain the restored image. This invention improves the structural consistency and detail fidelity of the restored image under non-rigid deformation through boundary-aware spiral sampling, non-folded alignment, and closed-loop feedback.