556 results about "Image alignment" patented technology

The invention relates to an image processing technology, and discloses an efficient panoramic image splicing method and system based on multi-view fusion, and the method comprises the steps: collecting a plurality of images from different views; performing multi-scale feature extraction on each image, generating a feature descriptor for each extracted feature point, and matching a corresponding feature point pair; performing multi-view geometric constraint screening on the feature point pairs; according to the screened feature point pairs, estimating a homography matrix between adjacent images, and carrying out global optimization on the homography matrix; aligning all the images into the same coordinate system; determining an overlapping region between adjacent images; and according to the pixel information in the overlapping areas, fusing the overlapping areas by adopting a self-adaptive weighted fusion algorithm so as to splice the plurality of images into a panoramic image. The invention further discloses a control device and a computer readable storage medium. The invention aims to improve the efficiency and accuracy of generating the multi-view fused panoramic image.

The invention discloses a multi-modal medical image fusion diagnosis system based on artificial intelligence, and the system comprises the following steps: extracting shared features of CT and MRI images through a convolutional neural network, and mapping the shared features to the same feature space; a two-way step-by-step alignment strategy is adopted, a three-dimensional deformation field matrix is generated, and cross-modal image anatomical structure alignment is achieved; calculating modal feature weights and eliminating distribution differences through an attention mechanism and an adversarial domain adaptation layer; constructing a CT-MRI image block contrast learning task, and optimizing a shared feature encoder; a conditional generative adversarial network is used for generating a false image of a missing mode according to the semantic segmentation map, and data distribution is constrained through a Wasserstein distance; uniform feature extraction of multi-modal medical images is realized through a shared feature encoder, the cross-modal image alignment accuracy is improved in combination with a bidirectional deformation field prediction module, and the comprehensiveness and accuracy of fusion features are enhanced by using a multi-modal feature fusion module.

The invention relates to the technical field of three-dimensional model reconstruction, and discloses a three-dimensional model reconstruction method based on point cloud data processing, which comprises the following steps: step 1, point cloud data acquisition and multi-modal fusion: acquiring point cloud data through a laser radar, an RGB camera and multi-view three-dimensional reconstruction to obtain a point cloud set, aligning the point cloud data with an RGB image to obtain a point cloud image set; establishing a mapping relation between point cloud coordinates and image pixel coordinates through an internal parameter matrix and an external parameter matrix of the camera to obtain aligned multi-modal point cloud data; 2, preprocessing the point cloud data: performing cleaning and downsampling processing on the multi-modal point cloud data obtained in the step 1; by combining point cloud data obtained by a laser radar, an RGB camera and multi-view three-dimensional reconstruction, accurate alignment of different sensor data is realized, high-precision and complete multi-modal point cloud data is obtained, blind areas of all sensors are effectively supplemented, and the precision and reliability of a three-dimensional reconstruction result are improved.

The invention provides a liver focus three-dimensional modeling method, and belongs to the technical field of image processing based on computer vision. Firstly, a multi-view spatial registration method based on optical flow optimization is designed, pixel-level displacement information of different view images is estimated by calculating an optical flow field, accurate image alignment is achieved, and spatial consistency of three-dimensional reconstruction is improved. And secondly, a three-dimensional reconstruction strategy based on two-dimensional focus segmentation is proposed, the two-dimensional focus segmentation is completed by adopting a lightweight U-Net variant, and a segmentation result is mapped to a three-dimensional space through a voxel probability projection method, so that 3D focus reconstruction is realized, and the calculation cost is reduced. And finally, extracting high-frequency features of the three-dimensional model by adopting a local edge enhancement method based on a Laplacian operator, and strengthening a focus boundary and a key anatomical structure through interpolation optimization, so that the three-dimensional model is more accurate and clearer. Compared with a traditional method, the method has the advantages that the mode of purely depending on image superposition is avoided, and the accuracy of three-dimensional modeling is improved.

The invention relates to the technical field of building structure health monitoring, in particular to a building defect detection method based on infrared thermal imaging and visible light bimodal fusion and a matched intelligent imaging device. According to the method, infrared thermal imaging and visible light bimodal image fusion are combined, and an improved YOLOv8m-seg model is combined, so that accurate detection of the defects of the outer wall of the building is realized. Firstly, a bimodal data set containing multi-material temperature difference data is constructed, ORB feature matching is combined with a feature alignment module (FAM) to achieve accurate image alignment, the defect detection robustness is improved through an improved bimodal feature fusion network, and finally automatic recognition and quantitative analysis of cracks, fractures and other defects are achieved. The problems that single-mode detection is interfered by dirt and is poor in multi-material adaptability are solved, the detection accuracy rate reaches 96% or above, the recall rate exceeds 93%, the single image detection time is shorter than or equal to 0.3 s, and the method is suitable for efficient detection of various building outer wall defects.

The invention relates to the technical field of computer vision, in particular to a reloading pedestrian re-identification method based on a visual language pre-training model. The method comprises the following steps: a training stage: inputting an image to obtain a clothing mask image, and generating clothing irrelevant / relevant prompts; text encoder parameters are fixed, prompt weights are optimized, text prompts are input into an encoder to obtain features, and a classifier is constructed to achieve image-text alignment through cross entropy loss; using a visual encoder to extract mask pattern features, and constraining a class center Euclidean distance to realize image-image alignment; stripping clothes characteristics: extracting clothes area characteristics and corresponding text characteristics, optimizing by a classifier, and introducing orthogonal loss to decouple clothes correlation; in the reasoning stage, a query image is input into a trained image encoder to extract features, and cosine similarity ranking and result returning are calculated according to the features of the image library. According to the technical scheme, the recognition accuracy of the pedestrian re-recognition method under the condition of pedestrian clothing change can be improved.

The invention relates to the technical field of image processing, and discloses an infrared image enhancement method and system based on local phase correlation, and the method comprises the steps: obtaining a plurality of continuous frames of infrared images, carrying out the intelligent partitioning of a reference frame, and calculating the variance feature, the method comprises the following steps: selecting regions of interest with rich information, independently executing phase correlation operation in each region to extract a local translation vector, obtaining global displacement estimation through weighted fusion, adopting an abnormal value detection algorithm to improve robustness, and finally realizing sub-pixel-level image alignment and intelligent weighted fusion. The method is suitable for real-time enhancement processing of satellite-borne infrared remote sensing images, the resource constraint requirement of an embedded platform is met while the processing quality is guaranteed, and an efficient and reliable technical scheme is provided for space remote sensing image processing.

The invention provides a biodiversity monitoring method and device and terminal equipment, and is suitable for the technical field of data processing, and the method comprises the steps: obtaining biological image information and biological sound information; performing feature extraction on the biological sound information to obtain multiple pieces of biological audio feature information; performing feature extraction on the biological image information to obtain multiple pieces of biological image feature information; performing alignment processing on the biological audio feature information and the biological image feature information to obtain a biological audio alignment feature and a biological image alignment feature; enhancing the biological audio alignment feature and the biological image alignment feature to obtain a biological audio monitoring feature and a biological image monitoring feature; and monitoring organisms according to the biological audio monitoring features and the biological image monitoring features. According to the invention, the audio information and the image information are combined, so that the accuracy and comprehensiveness of real-time monitoring of biodiversity are improved.

The invention discloses a task-oriented grabbing method and system based on cross-level constraint reasoning, and belongs to the technical field of robot grabbing control. The method comprises the following steps: aligning a text instruction with an input image through a mask alignment module, generating a target area mask by utilizing SAM-Clip, and generating a target point cloud by taking the target area mask as spatial prior; and further analyzing an internal physical structure of the target by using VLM, guiding to generate a 6-DOF grabbing attitude, performing collision detection and quality scoring by combining high-level function and bottom-level geometric prior, and outputting an optimal grabbing attitude. According to the method, the problems that task understanding and scene perception are disjointed, and the grabbing posture lacks constraint are solved, and the grabbing success rate and the intelligent level of the robot in the open environment are remarkably improved.

The invention discloses a tumor space-occupying brain network neural image alignment method based on multi-modal fusion, and belongs to the technical field of medical image processing and artificial intelligence crossing. The method comprises the following core steps of multi-modal image heterogeneous feature decoupling, tumor occupation deformation field modeling, functional network topological structure maintenance, cross-modal feature adversarial alignment, dynamic deformation constraint optimization and clinical interpretability verification, and construction of a three-dimensional non-rigid registration network based on a double attention mechanism. And differential homeomorphic mapping of a tumor focus area and normal brain tissue is realized through the cascaded spatial transformation module. Aiming at the problems of insufficient multi-modal feature alignment and brain network topology distortion in the prior art, the invention provides a function connection constrained cross-modal fusion strategy, a graph convolution network is adopted to encode resting state function connection features, and network node displacement caused by tumor occupation is dynamically corrected in combination with deformable convolution and a bidirectional feature competition mechanism; a space consistency loss function based on white matter fiber bundle tracing is designed, and through diffusion tensor imaging feature guide structure-function bimodal joint optimization, the problems of insufficient registration precision in a focus area and whole brain network connection distortion of a traditional method are solved. Experiments show that the registration precision of the method in glioma cases reaches 0.82 mm and is improved by 37% compared with that of a traditional method, and dissection-function consistency of functional network reconstruction around tumors is remarkably improved.

The invention discloses a house full life cycle intelligent monitoring method based on machine vision, particularly relates to the field of machine vision, and is used for solving the problems of insufficient image alignment precision, aging change recognition lag and incapability of accurately quantifying life cycle evolution trend in existing house structure aging detection. The method comprises the following steps of: acquiring an image sequence by setting a shooting point position of a house facade component, performing spatial composite alignment by utilizing affine transformation and feature edge fitting, extracting component key points in an aligned image, dividing the component key points into segmented regions, and constructing a structure life cycle image stacking group; then calculating color distribution dispersion and texture roughness entropy of each section of region, establishing an aging evolution feature sequence, executing trend mutation point detection, extracting a material state jump section, and constructing an aging trend jump map; and finally, based on the map, identifying a jump risk area and outputting house component aging information, thereby realizing accurate monitoring and trend study and judgment of the house outer wall structure state.

The invention relates to the technical field of intelligent robot path planning based on visual identification, and discloses an intelligent robot path planning method based on visual identification. According to the scheme, the problems of inconsistency of environment sensing data and lack of accurate geometric information are solved by utilizing data acquisition and calibration processing of the camera and the laser radar. The camera collects images, and parameters are determined by hardware; the internal reference matrix of the camera ensures clear correspondence between pixels and actual coordinates. The laser radar collects point cloud data, records three-dimensional coordinates of the environment, and is high in precision and real-time. Through data fusion and calibration, point cloud and image alignment by the system, camera internal reference matrix establishment and external reference calibration, accurate mapping and back projection are realized, and grassland area and obstacle information is extracted. According to the method, the sensing precision is improved, support is provided for path planning and navigation, and the robot has accurate positioning and efficient path planning capabilities.

The invention discloses a multi-frame photoacoustic image reconstruction method based on optical flow alignment and depth feature fusion. The method comprises the following steps: S1, obtaining photoacoustic signal data; s2, reconstructing a photoacoustic cross-sectional image; s3, performing optical flow calculation and image alignment; and S4, training the deep feature fusion network. According to the method, through optical flow motion correction and a potential space learning mechanism, space-time information and complementary features in the aligned multiple frames of images are dynamically integrated, complementary information in the multiple frames of images is adaptively fused, noise is suppressed, and finally reconstruction of high signal-to-noise ratio and high spatial resolution images of biological tissues is achieved. Experimental results show that the method can significantly improve image quality, recover image distortion and detail loss caused by motion and noise, and provide a new effective scheme for promoting robust clinical application of a photoacoustic imaging technology.

The invention provides a training-free text-image generation method based on a diffusion model, and relates to the technical field of computer graphic processing and artificial intelligence. The method comprises the following steps: extracting semantic phrases and layout information in an input text by utilizing a natural language model, inputting the input text, the semantic phrases and the layout information as additional conditions into a diffusion model, and extracting cross attention maps of different time steps; a positive and negative sample concept and a foreground and background concept based on an object are constructed, a new loss function is calculated on a cross attention map for semantic information and layout information, the loss function combines semantic loss, regional loss and original loss of a diffusion model and is used for updating a potential space image, and the image is generated through iterative denoising and a decoder. According to the method, additional training is not needed, image generation output based on the diffusion model better meets text requirements, and a better text and image alignment effect is achieved.

The invention discloses a brain multi-modal multi-sequence data registration method and device based on deep learning. The method comprises the steps of performing first iteration processing on a target image and a first moving image to obtain a first deformation field, wherein the first iteration processing comprises image alignment constraint processing of the target image, smooth constraint processing of the first deformation field, and area alignment constraint processing of a tumor area in the image; registering the first moving image to the target image based on the first deformation field to obtain a second moving image; performing second iteration processing on the target image and the second moving image to obtain a second deformation field; and registering the second moving image to the target image based on the second deformation field to obtain a registered moving image. In the process of performing unsupervised training on the deep learning model, after multiple times of iterative processing, model parameters are updated based on image alignment constraint loss with a target image, deformation field smooth constraint loss and region alignment constraint loss of a tumor region in the image.

The invention belongs to the field of characterization learning, and particularly relates to a prostate cancer diagnosis method based on a multimodal large model prompt learning mechanism. Comprising the following steps: step 1, data preprocessing; 2, key frames and similarity are calculated, and irrelevant editing is compressed; step 3, text and image alignment training; step 4, a test stage; according to the method, a similarity-based screening mechanism is provided, ultrasonic videos under coarse-grained labels are preliminarily screened under segmentation of a large model, and focus areas are focused on in time sequence; meanwhile, a pre-processing mechanism based on a large model is provided, and the influence of a large amount of irrelevant information existing in prostate ultrasonic image scanning on the model is compressed at a data end, so that the diagnosis effect of the model is improved.

The invention discloses a model training method, electronic equipment and a computer readable storage medium, and relates to the technical field of data processing and large models. The method comprises the steps of obtaining a training data set, wherein the training data set comprises image annotation data of various types of concerned objects and description text annotation data associated with the image annotation data; training the initial multi-modal retrieval model by adopting the training data set to generate a learnable prompt; carrying out image coding training on the initial multi-modal retrieval model by adopting the training data set and the learnable prompt to generate an intermediate multi-modal retrieval model; the training data set is adopted to conduct text and image alignment training on the middle multi-modal retrieval model, a target multi-modal retrieval model is generated, and the target multi-modal retrieval model is used for conducting multi-modal retrieval on the target retrieval request to obtain a target retrieval result. According to the method and the device, the technical problems that a multi-modal retrieval model in related technologies cannot adapt to variability and large-scale data requirements, and the model performance is relatively poor are solved.

The invention discloses an HDR video reconstruction method based on a standardized stream, and belongs to the technical field of high dynamic range image processing. The method comprises the following steps of: firstly, constructing a convolution optical flow estimation module with a self-adaptive normalized structure, wherein the convolution optical flow estimation module is used for accurately acquiring optical flow information between adjacent frames in an alternative exposure LDR video image sequence; then, carrying out multi-level feature alignment on the image sequence through an image alignment module so as to reduce alignment errors caused by illumination difference and movement; and finally, inputting the aligned and fused multi-level LDR image features into a standardized flow reconstruction network to realize high-quality HDR video image reconstruction. Aiming at the video reconstruction problem under the alternate exposure condition, the invention designs a standardized flow modeling structure considering the optical flow estimation precision and the feature alignment effect, and effectively improves the HDR video reconstruction quality in a complex dynamic scene.

The invention provides a multi-source perception method suitable for an unstructured scene of engineering machinery, and relates to the technical field of engineering machinery environment perception, and the method comprises the steps: carrying out the preprocessing of LiDAR point cloud data, generating a 2D perspective view aligned with an RGB image, and providing a basis for the subsequent feature extraction; semantic segmentation is carried out on image branches by using an improved deep learning network, and spatial depth features are extracted by point cloud branches through an optimized network structure. And the two are subjected to feature complementation through an FFAM module, so that the ability of understanding a complex scene is remarkably enhanced. Furthermore, a refinement module is introduced, an advanced 3D sparse convolution technology is adopted, and optimization processing is performed on point cloud features after back projection, so that the problem of fuzzy target boundary segmentation is effectively solved.

The invention relates to the technical field of image processing, and discloses a fluorescent lifetime microscopic image large-view-field splicing method and system and electronic equipment, and the method comprises the steps: constructing a tissue region mask for a to-be-spliced image, and constructing a vignetting model in a tissue region to complete vignetting correction; counting brightness indexes in an organization area to determine a global brightness reference, and realizing image group brightness unification through global zooming and single image brightness adaptive correction; further realizing alignment of adjacent images through feature point matching, and finishing image block splicing in combination with a minimum color difference suture fusion method to obtain an image band; an overlapping area is extracted from an image belt to generate an effective content mask, relative displacement is estimated by using a phase correlation method after low-frequency suppression, image belt splicing is completed by multiplexing a minimum color difference suture fusion method, a large-view-field spliced image is output, the automation degree and robustness of splicing are improved, the spliced image is geometrically consistent and visually seamless, and the splicing efficiency is improved. And the requirements of medical research on high-resolution and large-field-of-view fluorescence lifetime microscopic images are met.

The invention discloses a multi-modal fine-grained semantic alignment method and device based on a graph neural network, and relates to the field of multi-modal deep learning. Firstly, deep feature extraction is performed on input multi-modal original data, and then word-level text features and local image features are constructed into a cross-modal graph structure. And performing weighted aggregation on node neighborhood information of the cross-modal graph structure through the graph attention network. And finally, carrying out weighted fusion on the text alignment features and the image alignment features. According to the cross-modal feature fusion method, the word-level text features and the local image features are uniformly abstracted into the graph structure nodes for refined alignment, a more accurate cross-modal semantic corresponding relation can be captured, the heterogeneity problem in expression modes and semantic structures is effectively relieved, and the accuracy and reliability of cross-modal feature fusion are improved. The graph attention network can adaptively adjust the weight distribution of information propagation, highlights the effect of key features in the alignment process, and ensures that the model makes full use of important semantic relationships.

The invention provides a panoramic image intelligent splicing method based on context semantics. The method comprises the following steps: firstly, obtaining and preprocessing a sequence image; extracting a depth feature map of the image by using a pre-trained depth network; performing semantic-guided feature matching and image alignment based on the feature map; an optimal suture line is generated according to the semantic region boundary, and fusion is carried out by adopting a semantic weighted multi-band fusion algorithm; and finally, detecting and repairing the semantic inconsistent region to ensure the semantic coherence of the panoramic image. According to the method, matching is carried out by using the high-dimensional feature tensor rich in semantic information extracted by the deep neural network, the problem of feature matching ambiguity caused by repeated textures and weak texture regions such as sky or white walls is solved, the semantic information is used as a strong constraint, regions with similar appearances but different semantics can be effectively distinguished, and the accuracy of feature matching is improved. Therefore, correct matching point pairs with consistent semantics are screened out from massive candidates, and the matching accuracy is greatly improved.

The embodiment of the invention discloses an image alignment method, computer equipment and a computer storage medium. Taking the sub-region as a template image, taking an image obtained by externally expanding each edge of the sub-region by a preset size as an alignment image, and performing template matching on the template image and the alignment image according to a template matching algorithm to obtain a plurality of alignment pixel points and a correlation coefficient corresponding to each alignment pixel point; and according to the correlation coefficient of the plurality of alignment pixel points and the gray value of each alignment pixel point, determining the coordinate of a target alignment pixel point of the sub-region. And determining the alignment coordinates of the target die image according to the coordinates of the target alignment pixel points of the plurality of sub-regions. Since the alignment coordinate is calculated based on the pixel distribution condition and the periodic change condition of each sub-region of the target die image, the target die image and the reference image can be accurately aligned based on the alignment coordinate, the influence of the periodic pattern of the die image on the image alignment can be avoided, and the alignment precision of the die image is improved.

The embodiment of the invention provides a tunnel disease detection method and system based on an unmanned aerial vehicle, and belongs to the technical field of defect optical detection. The method comprises the steps that an unmanned aerial vehicle is controlled to fly along a tunnel to collect multichannel image data of the surface of a structure, and the flight attitude is adjusted based on environment illumination information to execute image illumination compensation; performing image alignment of each target anchor point based on a structure anchor point atlas constructed based on historical acquisition images in combination with the flight pose information of the unmanned aerial vehicle and the multi-channel image data after illumination compensation; identifying a disease area in the aligned image, extracting disease features of each target anchor point at this time, and updating the disease of each target anchor point at this time into a time sequence feature data sequence corresponding to each target anchor point; and based on the updated time sequence characteristic data sequence of each target anchor point, Bayesian point change detection is adopted to analyze the disease evolution trend of the tunnel. According to the scheme, the alignment precision, the time sequence comparability and the risk judgment capability of tunnel disease detection are integrally improved.

The invention discloses an SPR (Surface Plasmon Resonance) image optimization processing method based on image segmentation and edge enhancement, which comprises the following steps: acquiring SPR image data, and preprocessing to generate standardized SPR image data; inputting a structure boundary extraction model constructed based on CGNet, generating a structure boundary label graph, and aligning the structure boundary label graph with the image; gradient amplitude and local entropy mutation detection is executed, and an artifact guide graph is generated; respectively inputting the standardized image into details and context branches of the improved CSDNet, and extracting edge and semantic feature maps; inputting a guide perception gating module, executing structure enhancement and artifact suppression fusion, and generating a fusion feature map; inputting into a multi-scale detail recovery module, and outputting an edge enhanced image; and executing structural similarity and marginal definition scoring based on the original image and the enhanced image, and generating an optimization result. According to the method, synchronous optimization of SPR image edge enhancement and structure maintenance is realized, and the image definition and diagnosis availability are remarkably improved.

The invention discloses a plastic mold injection molding defect detection method and system, and the method comprises the following steps: collecting a surface image corresponding to a plastic mold injection molding part and an internal tomography image, carrying out the denoising, graying and image alignment, and meanwhile, generating a multi-dimensional image data set corresponding to each injection molding part through an associated index; extracting geometric features and texture features corresponding to each injection molding part based on the multi-dimensional image data set, and fusing the geometric features and the texture features to generate geometric texture vectors; performing similarity judgment estimation based on the geometric texture vector corresponding to each injection molding part and a preset defect-free injection molding part feature template, recursively determining a suspected defect area of each injection molding part, and estimating and generating an injection molding part defect type parameter table; and analyzing and generating injection molding part process defect optimization suggestions based on the injection molding part defect parameter table, and meanwhile, carrying out associated storage to generate an injection molding defect detection archive library. According to the invention, the mold cavity position corresponding to the defect can be positioned, and the mold injection molding defect detection efficiency can be greatly improved.

The invention discloses an ANA detection image-text report intelligent generation method and system, and relates to the technical field of immunodetection, the system uses different fluorescent markers to mark ANA, signal sensitivity is enhanced, and detection accuracy is ensured. Then, collecting fluorescence signals of different wavelengths by adopting a fluorescence microscope and a multi-wavelength excitation light source, and obtaining dynamic change data of reaction intensity by combining a time-resolved fluorescence imaging technology to generate a multi-channel imaging data set; thirdly, noise removal, signal enhancement and image alignment are carried out through an image preprocessing technology, and data consistency is ensured; based on a multi-channel data set, a convolutional neural network is used for training and testing an initial model, an intelligent detection model is constructed, a fluorescence reaction mode is extracted, finally, a detailed ANA detection report is automatically generated by calculating a comprehensive scoring coefficient and comparing the comprehensive scoring coefficient with a preset threshold value, and an accurate immunodiagnosis basis is provided for doctors.

The embodiment of the invention aims to provide a method and equipment for carrying out empty ear identification on rice ears based on image analysis and a computer program product. The method comprises the following steps: acquiring an RGB image and an NIR image of a single rice spike; aligning the RGB image and the NIR image and converting the RGB image and the NIR image into corresponding four-channel image data; and inputting the four-channel image data into a double-path classification model to obtain an empty ear identification result output by the double-path classification model. Each embodiment of the invention provides a double-path empty ear recognition scheme, and the industrial problem of coexistence of artificial experience limitation and deep learning overfitting in a traditional method is effectively solved through a dynamic complementation mechanism of preset classification features and deep image features. The two types of features are subjected to cross-modal interaction of classification feature contribution degree weighting, and when explicit color features are abnormal, the decision proportion of deep texture features is enhanced, so that the overall detection accuracy is kept stable.

The invention discloses a hierarchical three-dimensional scene generation method and system based on spatial super-division, and belongs to the technical field of computer graphics, and the method comprises the steps: carrying out the preprocessing of a scene image, and obtaining a high-resolution object image; generating initial rough scene voxels for the scene image, and screening rough voxels and structural latent variables aligned with the high-resolution object image from the initial rough scene voxels to construct a hierarchical scene tree; inputting the high-resolution object image and the rough voxel into a voxel super-resolution model, and generating a fine voxel which keeps geometric consistency with the rough voxel; performing scale alignment and attitude registration based on the rough voxels and the fine voxels; and generating fine voxels of the sub-components recursively by taking the rough voxels of the current node as conditions based on the hierarchical scene tree, and finally assembling to generate a high-resolution three-dimensional scene. According to the method, a high-quality three-dimensional scene with high visual fidelity, fine geometric details and global structure consistency can be efficiently and automatically reconstructed from a single RGB image.

The embodiments of this application provide an image data collection system, an image model training method, and a device for improving image resolution. In this application, an image capturing device is used to capture images of an object at different focal lengths to obtain a first image and a second image respectively, and the first image and the second image are processed to obtain a first processed image with high resolution and a second processed image with low resolution, respectively. Image alignment is performed on these processed images to obtain a high-resolution and low-resolution image pair. Many high-resolution and low-resolution image pairs are collected as a training image dataset to train a model for upgrading low-resolution images to high-resolution images. The trained model can significantly improve the ability to restore image details.