CN116503552A - Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method - Google Patents

Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method Download PDF

Info

Publication number
CN116503552A
CN116503552A CN202310404149.5A CN202310404149A CN116503552A CN 116503552 A CN116503552 A CN 116503552A CN 202310404149 A CN202310404149 A CN 202310404149A CN 116503552 A CN116503552 A CN 116503552A
Authority
CN
China
Prior art keywords
point cloud
coarse
features
local
fine
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310404149.5A
Other languages
Chinese (zh)
Inventor
张德军
王杨
徐战亚
吴亦奇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China University of Geosciences
Original Assignee
China University of Geosciences
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China University of Geosciences filed Critical China University of Geosciences
Priority to CN202310404149.5A priority Critical patent/CN116503552A/en
Publication of CN116503552A publication Critical patent/CN116503552A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Multimedia (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Geometry (AREA)
  • Computer Graphics (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention provides a method for supplementing a shape of a coarse-to-fine point cloud based on multi-scale feature fusion, which comprises the following steps: stage 1, generating a point cloud skeleton: extracting global features from incomplete point cloud XSupplementing a rough point cloud Y through a multi-layer perceptron coarse Taking the three points as a skeleton of the complete point cloud; stage 2, refining geometrical features: skeleton Y using a complete point cloud coarse And the incomplete point cloud X learns to multi-scale local geometric featuresFusing global featuresAnd local geometric featuresThe fused features are gradually up-sampled and a fine complete point cloud Y' is generated through a multi-layer perceptron. The beneficial effects of the invention are as follows: the advantages of different characteristics are effectively utilized, and the accuracy of the point cloud complement neural network is improved; and the neural network is rolled based on the deep learning chart, so that the accuracy of the point cloud complement result is improved.

Description

Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method
Technical Field
The invention relates to the field of point clouds, in particular to a method for supplementing the shape of a coarse-to-fine point cloud based on multi-scale feature fusion.
Background
The traditional three-dimensional shape complement method usually adopts voxels or grids to describe a three-dimensional object, and then three-dimensional convolution is used for extracting features, so that great success is achieved in tasks such as shape complement and three-dimensional reconstruction. However, these methods suffer from the same drawbacks that voxel-based methods consume a lot of memory and cannot generate high-resolution three-dimensional shapes.
To increase resolution, it is proposed to progressively voxel specific regions using octree structures. Recent years have focused on three-dimensional reconstruction of gridding due to the quantization effect of the voxelization operations. The existing mesh representation is a target mesh deformed based on a template mesh, and is only valid for a specific type of three-dimensional data. Thus, researchers turn their eyes towards using unstructured point cloud data as a representation of three-dimensional objects, which results in less computational memory consumption due to their unique data representation, but with the same data size, the fine-grained representation of three-dimensional shapes is more powerful.
However, three-dimensional convolution is not suitable for processing point cloud data because regular convolution operators are not suitable for computing unordered data types. PointNet and PointNet++ are pioneering efforts to directly process three-dimensional point clouds, providing new ideas for point cloud data processing and inspiring research for many downstream tasks.
The point cloud data is extremely easy to acquire, and has the advantages that the storage is convenient, the calculation cost required by processing is lower than that of voxel data, and the like. In recent years, shape complement tasks for point cloud data have been attracting attention. The point cloud complement aims to generate a complete three-dimensional point cloud shape according to the locally observed point cloud. PCN, topNet, SA-Net, ECG is a framework that uses encoder-decoder to achieve point cloud completion. Most of these methods use a PointNet or PointNet++ encoder to extract global features from the incomplete input point cloud and use a decoder to infer the complete point cloud shape from the extracted global features. However, the point cloud complement methods all have a common problem: they do not fully exploit the local information of the point cloud, resulting in a lack of clear local detail of the generated point cloud.
The PCN firstly learns global features from the incomplete point cloud, then generates a coarse complete point cloud, and finally uses the floating-Net to up-sample the complete point cloud to obtain a fine complete point cloud. TopNet predicts the complete point cloud shape using a tree structured decoder. SA-Net adopts a multi-stage Folding structure and skip-coating to repair incomplete point cloud shapes. The ECG firstly uses global features to generate a rough point cloud shape, and then uses edge local features to generate a fine point cloud shape.
To preserve and recover local geometric detail, the local features are utilized to refine the coarse complete point cloud. NSFA combines known features and predicted missing features to complement the complete three-dimensional point cloud shape. NSFA assumes a ratio of known to missing parts of about 1:1 (the observable part is about half of the whole object), however this a priori condition is in most cases very demanding. VRCNet focuses on the framework of an anti-learning and variational automatic encoder to improve the authenticity and consistency of the generated complete shape. Xie et al propose a point cloud generator based on style and contrast differentiable rendering for point cloud completion. Yu et al convert point cloud completion into set-to-set conversion problems and propose a Transformer encoder-decoder architecture (PoinTr) that can better learn structural knowledge, preserve detailed information to complete point cloud completion. However, the network structure of these methods is complex, the network is easy to be over-fitted, and some prior information is needed as a premise.
Disclosure of Invention
In order to solve the above problems, the present invention provides a coarse-to-fine point cloud shape complement method (Multi-scale Feature Fusion Network, MFF-Net) based on Multi-scale feature fusion, mainly comprising:
stage 1, generating a point cloud skeleton:
extracting global features from incomplete point cloud X
Supplementing a rough point cloud Y through a multi-layer perceptron coarse Taking the three points as a skeleton of the complete point cloud;
stage 2, refining geometrical features:
skeleton Y using a complete point cloud coarse And the incomplete point cloud X learns to multi-scale local geometric features
Fusing global featuresAnd local geometry->
The fused features are gradually up-sampled and a fine complete point cloud Y' is generated through a multi-layer perceptron.
Further, SA of PointNet++ is adopted to extract global structure information from incomplete point cloud X
Further, the SA gradually abstracts larger and larger local areas in a multi-level hierarchical structure, and more local neighborhood features can be focused when the features of the points with different resolutions are acquired; the multi-level hierarchical encoder is composed of a series of SAs; the SA includes a sampling layer, a grouping layer, and a PointNet layer.
Further, the sampling layer samples the input incomplete point cloud with the most distant point to select a group of points as the mass centers of the respective local neighborhoods, the grouping layer constructs a local neighborhood set by searching adjacent points around the mass centers, and the PointNet layer encodes the local neighborhood points as feature vectors by using MLPs.
Further, the global features are decoded by a multi-layer perceptron to generate a rough complement result, namely a rough point cloud Y coarse As a skeleton of the complete point cloud.
Further, the method adopts a global and layout feature fusion module with a soft attention mechanism to the global featuresAnd local geometry->Fusion is performed.
Further, after fusion, the information of K adjacent points around each point is aggregated by using a DB with the function of a graph network, and then the information is cascaded with the original point characteristics to form new local characteristics, and then the acquired local characteristics are downsampled by an EP module.
Further, a K-nearest neighbor algorithm is used to select K nearest neighbor points in the feature space for each point to construct a neighborhood graph, thereby extracting edge features.
The technical scheme provided by the invention has the beneficial effects that: the advantages of different characteristics are effectively utilized, and the accuracy of the point cloud complement neural network is improved; and the neural network is rolled based on the deep learning chart, so that the accuracy of the point cloud complement result is improved.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a flow chart of a method for shape complement from coarse to fine point cloud based on multi-scale feature fusion in an embodiment of the invention.
Detailed Description
For a clearer understanding of technical features, objects and effects of the present invention, a detailed description of embodiments of the present invention will be made with reference to the accompanying drawings.
The embodiment of the invention provides a method for supplementing the shape of a coarse-to-fine point cloud based on multi-scale feature fusion. And respectively defining the point cloud Y and the point cloud X as a complete three-dimensional point cloud object and a locally observed incomplete point cloud thereof. Wherein X is not a subset of Y. The point cloud completion task is to predict a complete three-dimensional point cloud shape Y' under the condition of a given point cloud X. Samples of X and Y can be easily obtained using a large-scale synthetic dataset, which is implemented using supervised learning methods.
Referring to fig. 1, fig. 1 is a flowchart of a method for completing a shape of a coarse-to-fine point cloud based on multi-scale feature fusion according to an embodiment of the present invention, which specifically includes:
the method uses subtasks of two stages to realize point cloud completion, wherein the two stages are respectively used for generating a point cloud skeleton and refining geometric details. In particular, the MMF-Net is made up of two sub-networks of encoder-decoders and gradually complements the point cloud shape Y' in a coarse-to-fine manner.
Stage 1: and generating a point cloud skeleton. Firstly, extracting global features by using a PointNet++ multi-level point Set extraction module (Set extraction, SA)Then complement a rough point cloud Y by a multi-layer perceptron (multilayer perceptrons, MLPs) coarse As a skeleton of the complete point cloud.
Stage 2: refining the geometric details. First using a coarse skeleton Y coarse And the incomplete point cloud X learns to multi-scale local geometric featuresECG fusion->And->The correlation between these two more complementary information is not explicitly considered in the process of (a). In order to combine the local geometric features intimately->And global features->The method designs a global and layout feature fusion module (global and local Feature, GLF) with a soft attention mechanism, while retaining the respective information. And finally, gradually upsampling the fused features and generating a fine complete point cloud Y' through a multi-layer perceptron.
The point cloud completion is realized in a rough-to-fine mode, the learned global features are decoded to generate a rough completion result which is used as a skeleton of the complete point cloud, and then the local geometric features are gradually recovered. Predicting the shape skeleton before refining the local geometry is necessary for the following reasons: (1) The shape skeleton describes a complete point cloud structure, i.e. a region missing in local observation is restored; (2) The shape skeleton may be considered as an adaptive three-dimensional anchor for mining local geometric features in the residual point cloud.
The goal of stage 1 codecs is to generate a complete point cloud structure, the main challenge of which is how to recover the complete three-dimensional shape from the incomplete point cloud while preserving certain local geometry. The method uses SA in PointNet++, gradually abstracts larger and larger local areas in a multi-level hierarchical structure, and can pay attention to more local neighborhood characteristics when acquiring the characteristics of points with different resolutions.
The multi-level hierarchical encoder is composed of a series of SAs. The SA includes a sampling layer, a grouping layer, and a PointNet layer. The sampling layer samples the input incomplete point cloud with the furthest point to select a group of points as the mass centers of the respective local neighborhoods. The grouping layer constructs a local neighborhood set by finding neighboring points around the centroid. The PointNet layer encodes local neighborhood points (including centroids) as feature vectors using MLPs. The encoder finally converts the input incomplete point cloud into a 1024-dimensional global feature, and then generates a rough complete point cloud through a decoder composed of MLPs.
The goal of the stage 2 codec is to further generate a high quality full point cloud shape based on the observed incomplete point cloud and the complement coarse point cloud shape. However, in incomplete, unevenly distributed, and sparsely isolated point clouds, efficient extraction of edge geometry is challenging. Thus, the global shape structure of the coarse point cloud X and the local geometric details in the incomplete point cloud X are utilized to avoid this problem. The point cloud shape skeleton may be considered an adaptive three-dimensional anchor point that facilitates capturing useful local edge features from the point cloud of flaws.
Both stage 1 and stage 2 employ a codec comprising DB, EP and GLF modules.
As shown in FIG. 1, the input of stage 2 includes a complement of coarse point cloud Y coarse And the original incomplete point cloud X is cascaded, and then the structure skeleton of the complete point cloud is provided and the local geometric details of the input point cloud are reserved. And then, using DB (dense block) with graph network function to aggregate the information of K adjacent points around each point, cascading with the original point characteristics to form new local characteristics, and downsampling the acquired local characteristics through an EP (edge-preserved pooling) module. The above process aggregates the individual point features of each point neighborhood K, resulting in lower resolution point features while preserving more characterizable local features. The K nearest neighbor algorithm is used to select K nearest neighbor points in the feature space for each point to construct a neighborhood graph, thereby extracting edge features. Edge preservation pooling is employed. When the input points are resampled, the characteristics of the selected centroid point and the local neighborhood points thereof are simultaneously propagated to the subsequent network layers for processing. Thus, as the point set becomes more sparse, it experiencesThe field becomes larger. To learn multi-scale edge features of points more efficiently. Next, the local features generated by aggregation are feature-fused with global features of the same resolution generated by the 1 st stage encoder-decoder using GLF, and the fused features are used as input of the next DB. The feature fusion module is used for complementing the advantages of the global features and the local features and enhancing the network capability. The above process is repeated to obtain the characteristics of the multi-resolution points. Finally, the multi-resolution features generated by the encoder-decoder of the 2 nd stage and the features obtained by gradually up-sampling (US) the global features are respectively cascaded, and fine point clouds are generated through MLPs.
The beneficial effects of the invention are as follows: the advantages of different characteristics are effectively utilized, and the accuracy of the point cloud complement neural network is improved; and the neural network is rolled based on the deep learning chart, so that the accuracy of the point cloud complement result is improved.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims (8)

1. A method for supplementing the shape of a coarse-to-fine point cloud based on multi-scale feature fusion is characterized by comprising the following steps of: comprising the following steps:
stage 1, generating a point cloud skeleton:
extracting global feature F from incomplete point cloud X X G ；
Supplementing a rough point cloud Y through a multi-layer perceptron coarse Taking the three points as a skeleton of the complete point cloud;
stage 2, refining geometrical features:
skeleton Y using a complete point cloud coarse And the incomplete point cloud X learns to obtain multi-scale local geometric characteristics F X P ；
Fusing global features F X G And local geometric feature F X P ；
The fused features are gradually up-sampled and a fine complete point cloud Y' is generated through a multi-layer perceptron.
2. The method for completing the shape of the point cloud from coarse to fine based on multi-scale feature fusion according to claim 1, wherein the method comprises the following steps of: SA extraction of global structure information F from incomplete cloud X by using PointNet++, and method for extracting global structure information F from incomplete cloud X by using SA X G 。
3. The method for supplementing the shape of the point cloud from coarse to fine based on multi-scale feature fusion as claimed in claim 2, wherein the method comprises the following steps of: the SA gradually abstracts larger and larger local areas in a multi-level hierarchical structure, and can pay attention to more local neighborhood characteristics when acquiring the characteristics of different resolution points; the multi-level hierarchical encoder is composed of a series of SAs that include a sampling layer, a grouping layer, and a PointNet layer.
4. A coarse-to-fine point cloud shape completion method based on multi-scale feature fusion as claimed in claim 3, wherein: the sampling layer samples the input incomplete point cloud with the most distant point to select a group of points as the mass centers of the respective local neighborhoods, the grouping layer constructs a local neighborhood set by searching adjacent points around the mass centers, and the PointNet layer encodes the local neighborhood points into feature vectors by using MLPs.
5. The method for completing the shape of the point cloud from coarse to fine based on multi-scale feature fusion according to claim 1, wherein the method comprises the following steps of: decoding the global features by a multi-layer perceptron to generate a rough complement result, namely a rough point cloud Y coarse As a skeleton of the complete point cloud.
6. The method for completing the shape of the point cloud from coarse to fine based on multi-scale feature fusion according to claim 1, wherein the method comprises the following steps of: the method adopts a global and layout feature fusion module with a soft attention mechanism to realize global feature F X G And local geometric feature F X P Fusion is performed.
7. The method for completing the shape of the point cloud from coarse to fine based on multi-scale feature fusion according to claim 1, wherein the method comprises the following steps of: after fusion, the DB with the graph network function is used for aggregating the information of K nearest neighbor points around each point, and then the information is cascaded with the original point characteristics to form new local characteristics, and the acquired local characteristics are downsampled through the EP module.
8. The method for completing the shape of the point cloud from coarse to fine based on multi-scale feature fusion according to claim 7, wherein the method comprises the following steps of: the K nearest neighbor algorithm is used to select K nearest neighbor points in the feature space for each point to construct a neighborhood graph, thereby extracting edge features.
CN202310404149.5A 2023-04-14 2023-04-14 Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method Pending CN116503552A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310404149.5A CN116503552A (en) 2023-04-14 2023-04-14 Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310404149.5A CN116503552A (en) 2023-04-14 2023-04-14 Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method

Publications (1)

Publication Number Publication Date
CN116503552A true CN116503552A (en) 2023-07-28

Family

ID=87321019

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310404149.5A Pending CN116503552A (en) 2023-04-14 2023-04-14 Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method

Country Status (1)

Country Link
CN (1) CN116503552A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274764A (en) * 2023-11-22 2023-12-22 南京邮电大学 Multi-mode feature fusion three-dimensional point cloud completion method
CN117274454A (en) * 2023-08-29 2023-12-22 西交利物浦大学 Three-dimensional point cloud completion method, device and storage medium based on component information
CN117422645A (en) * 2023-11-14 2024-01-19 中国科学院长春光学精密机械与物理研究所 Confidence aggregation-based radar point cloud shape completion method
CN117475107A (en) * 2023-08-29 2024-01-30 北京航空航天大学 Relational enhancement point cloud completion method based on deep learning

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274454A (en) * 2023-08-29 2023-12-22 西交利物浦大学 Three-dimensional point cloud completion method, device and storage medium based on component information
CN117475107A (en) * 2023-08-29 2024-01-30 北京航空航天大学 Relational enhancement point cloud completion method based on deep learning
CN117422645A (en) * 2023-11-14 2024-01-19 中国科学院长春光学精密机械与物理研究所 Confidence aggregation-based radar point cloud shape completion method
CN117274764A (en) * 2023-11-22 2023-12-22 南京邮电大学 Multi-mode feature fusion three-dimensional point cloud completion method
CN117274764B (en) * 2023-11-22 2024-02-13 南京邮电大学 Multi-mode feature fusion three-dimensional point cloud completion method

Similar Documents

Publication Publication Date Title
Li Research and application of deep learning in image recognition
Lateef et al. Survey on semantic segmentation using deep learning techniques
CN116503552A (en) Multi-scale feature fusion-based coarse-to-fine point cloud shape completion method
US11301965B2 (en) Method and image processing device for image super resolution, image enhancement, and convolutional neural network model training
CN110322495A (en) A kind of scene text dividing method based on Weakly supervised deep learning
CN113362242B (en) Image restoration method based on multi-feature fusion network
Liu et al. Crackformer network for pavement crack segmentation
CN114241274A (en) Small target detection method based on super-resolution multi-scale feature fusion
CN115393289A (en) Tumor image semi-supervised segmentation method based on integrated cross pseudo label
Wang et al. TF-SOD: a novel transformer framework for salient object detection
CN115936992A (en) Garbage image super-resolution method and system of lightweight transform
Yu et al. WaterHRNet: A multibranch hierarchical attentive network for water body extraction with remote sensing images
Cheng et al. A survey on image semantic segmentation using deep learning techniques
Zhang et al. SegNet Network Architecture for Deep Learning Image Segmentation and Its Integrated Applications and Prospects
CN118134952A (en) Medical image segmentation method based on feature interaction
Chan et al. Asymmetric cascade fusion network for building extraction
Mu et al. Integration of gradient guidance and edge enhancement into super‐resolution for small object detection in aerial images
Tan et al. Exploring the Application of Large-Scale Pre-Trained Models on Adverse Weather Removal
Zheng et al. Transformer-based hierarchical dynamic decoders for salient object detection
Zheng et al. Dcu-net: Self-supervised monocular depth estimation based on densely connected u-shaped convolutional neural networks
Ray et al. Image segmentation and classification using deep learning
Zhu et al. Gradient-based graph attention for scene text image super-resolution
Wang et al. Feature re-attention and multi-layer feature fusion for fine-grained visual classification
Xu et al. A review of semantic segmentation based on context information
US11798197B2 (en) Data compression with a multi-scale autoencoder

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination