CN112669230B - Point cloud data denoising method based on convolutional neural network - Google Patents
Point cloud data denoising method based on convolutional neural network Download PDFInfo
- Publication number
- CN112669230B CN112669230B CN202011538245.1A CN202011538245A CN112669230B CN 112669230 B CN112669230 B CN 112669230B CN 202011538245 A CN202011538245 A CN 202011538245A CN 112669230 B CN112669230 B CN 112669230B
- Authority
- CN
- China
- Prior art keywords
- neural network
- data
- point cloud
- convolutional neural
- data block
- 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.)
- Active
Links
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a point cloud data denoising method based on a convolutional neural network, which comprises the steps of firstly obtaining original three-dimensional point cloud data, preprocessing the original data, screening abnormal values, normalizing a data format, constructing a convolutional neural network model, setting hyper-parameters of the convolutional neural network model A through functions in skearn to build the convolutional neural network model A, Bi, preparing training data, respectively training the convolutional neural network model A, Bi (i is 1-n), denoising by using the convolutional neural network model, and finally restoring denoised data by using a query algorithm. The method can effectively finish the denoising of the noise-containing point cloud data acquired by the three-dimensional scanning of the oil platform.
Description
Technical Field
The invention belongs to a method for processing noise contained in point cloud data, and particularly relates to a point cloud data denoising method obtained by three-dimensional scanning of an oil platform.
Background
In the process of exploiting and developing deep water resources in south China sea, large-scale drilling platforms are needed. In the construction of large-scale oil and natural gas exploitation equipment, a large-scale ocean engineering module is constructed in a sectional folding mode. In the process of layered construction, the precision of the construction and installation process needs to be ensured. In order to ensure the accuracy of installation, the control is generally performed from two aspects: the manufacturing precision of each layer is strictly controlled, and the mounting position of the structure is monitored by adopting a high-precision measurement method, so that the integral mounting precision is ensured. At present, no mature and complete installation precision control technology is developed for the measurement technology of large ocean engineering modules in China. Meanwhile, in the process of acquiring the three-dimensional point cloud data of the large-scale offshore engineering oil platform by using the three-dimensional laser scanning technology, the three-dimensional point cloud data of the large-scale offshore engineering oil platform obtained by scanning inevitably has noise due to errors of sensor measurement and uncertainty factors possibly contained in the platform, the implementation of a high-precision control technology is greatly influenced by the existence of the noise, the existing denoising algorithm mostly depends on the prior distribution of the point cloud data, the distribution usually needs a very high professional level, and the prior distribution of the noise possibly has differences aiming at different oil platforms, if the noise in the three-dimensional point cloud data obtained by a new oil platform through the three-dimensional laser scanning technology is denoised according to the traditional denoising method, the noise in the obtained three-dimensional point cloud data needs to be evaluated again, the prior distribution of the point cloud data is obtained, and the defects cause that the denoising process of the three-dimensional point cloud data obtained by the three-dimensional laser scanning technology is very complicated and has no mobility in the existing method.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a novel point cloud data denoising method based on a convolutional neural network for denoising noisy point cloud data.
In order to achieve the purpose, the invention adopts the technical scheme that:
the invention discloses a point cloud data denoising method based on a convolutional neural network, which comprises the following steps of:
the method comprises the following steps of firstly, acquiring original three-dimensional point cloud data: scanning a first oil-gas platform needing to obtain a three-dimensional configuration in an actual project by adopting a three-dimensional laser scanning technology, obtaining three-dimensional point cloud data of the first oil-gas platform as point cloud data to be denoised and geometrical space information of the point cloud data to be denoised, and simultaneously scanning a second oil-gas platform as a sample object by using three-dimensional laser to obtain three-dimensional point cloud data of the second oil-gas platform as point cloud data for a subsequent training model and obtain geometrical space information of the point cloud data for the subsequent training model;
secondly, preprocessing the original data, comprising the following steps:
first step, outlier screening: by outlier analysis, respectively removing outliers in the point cloud data to be denoised and the point cloud data of a subsequent training model by using the geometric space information of the three-dimensional point cloud data;
and step two, standardizing a data format, which comprises the following specific steps: sequentially cutting point cloud data to be denoised and point cloud data of a subsequent training model in a scanning direction to obtain a plurality of standardized point cloud data block matrixes to be denoised and point cloud data block matrixes of the subsequent training model, wherein each data block matrix is A, B and C, A represents the number of data contained in the length direction of a data block, B represents the number of data contained in the width direction of the data block, C represents the number of data contained in the height direction of the data block, two adjacent point cloud data block matrixes have a superposition area, the size of the superposition area is determined by a set superposition parameter alpha, and the size of the superposition area is (A, alpha) B, C;
step three, constructing a convolutional neural network model, and specifically comprising the following processes:
step one, setting hyper-parameters of a convolutional neural network model A through a function in sklern to build the convolutional neural network model A, wherein the hyper-parameters comprise: the method comprises the steps that the size of an input layer of a convolutional neural network model A is set to be the size of a standardized data block square matrix obtained through block cutting, namely A, B, C, a convolutional layer and a pooling layer are arranged in the convolutional neural network in an alternating structure, the depth of the convolutional layer and the pooling layer is set to be 10, each convolutional layer uses 64 convolutional kernels, the activation functions of the convolutional layer and an output layer select relu activation functions, square error loss is used as a loss function, a learning rate is automatically adjusted in a mode of adapting to the learning rate according to the training, finally the output layer is various Gaussian noise types contained in the data block square matrix with the size of A, B and C after the block cutting is extracted by the convolutional neural network, characteristic numbers n represent different types of Gaussian noise, and i is 1,2 and 3 … … … n;
secondly, adopting the method of the first step to re-establish n convolutional neural network models B with the same structure as the convolutional neural network model B in the first stepiI is 1 to n, and finally each convolution neural network model BiEach convolution neural network model B is a denoising model obtained by adding Gaussian noise of the same characteristic number category as the value i to a data block square matrix of an input layer and trainingiThe output result of the output layer is a data block square matrix with the Gaussian noise category corresponding to the characteristic number removed;
step four, training a convolutional neural network model, and the specific process is as follows:
the first step, training data preparation:
performing data enhancement on the point cloud data block square matrix of the subsequent training model, and simultaneously independently superposing different levels of Gaussian noise represented by each characteristic number to all data block square matrices obtained after the point cloud data of the subsequent training model is diced and standardized;
secondly, training a convolutional neural network model A, wherein the specific method comprises the following steps: inputting a convolutional neural network model A by using all the data block matrixes obtained in the first step after data enhancement and noise superposition, and training the convolutional neural network A by taking the noise category contained in the input data block matrixes as output to obtain a trained convolutional neural network A;
thirdly, training each convolution neural network model BiThe specific method comprises the following steps:
firstly, respectively superposing certain class of Gaussian noise in all data block matrixes of subsequent training models, and then respectively inputting the Gaussian noise into a convolution neural network model B with the same i value as the characteristic number of the Gaussian noise of the classiIn the method, all data block matrixes of the subsequent training model before the Gaussian noise is superposed are taken as each convolutional neural network model BiTo the convolutional neural network model BiTraining to obtain a trained convolutional neural network Bi；
And step two, continuously repeating the step one until all the convolutional neural network models B are completediTraining;
and fifthly, denoising by using the convolutional neural network model, wherein the specific process is as follows:
firstly, inputting all standardized square matrix data of point cloud data to be denoised into a trained convolutional neural network model A respectively, and classifying Gaussian noise contained in the standardized square matrix data of the point cloud data to be denoised by using the convolutional neural network A;
secondly, respectively inputting the standardized data block matrixes of all point cloud data to be denoised with the same characteristic number of noise category classification into a trained convolutional neural network model B with the same value of i as the characteristic number of Gaussian noise of the categoryiMiddle, convolution neural network model BiAfter outputting and removing the noise corresponding to the characteristic number of the noise class classificationAll data block matrixes of (1);
and step six, restoring the de-noised data, wherein the specific method comprises the following steps:
the method comprises the steps of firstly, randomly selecting a denoised first data block square matrix;
secondly, matching a second data block square matrix which is overlapped with the edge data of the first data block square matrix by using a query algorithm;
thirdly, splicing the data overlapped part of the first data block square matrix and the second data block square matrix;
and fourthly, repeating the second step and the third step, continuously splicing the adjacent data block matrixes of the current data block matrix by using a cyclic algorithm, and finally splicing all the data block matrixes to obtain point cloud data which is denoised and has the same structure as the input original noisy data.
The invention has the beneficial effects that: the method can effectively finish the denoising of the noisy point cloud data obtained by the three-dimensional scanning of the oil platform, and can realize the denoising of the noisy three-dimensional point cloud data without depending on the prior estimation of experts on the noise level in the scanned data. The adopted denoising algorithm is different from the dependence of the traditional denoising method on prior distribution, and the method can perform self-adaptive noise classification and denoising on the noisy point cloud data with unknown distribution, so that the robustness of a trained model is ensured, and the sensitivity to noise under different conditions is low.
Detailed Description
The present invention will be further described with reference to the following embodiments.
The invention discloses a point cloud data denoising method based on a convolutional neural network, which comprises the following steps of:
the method comprises the following steps of firstly, acquiring original three-dimensional point cloud data: scanning a first oil-gas platform needing to obtain a three-dimensional configuration in an actual project by adopting a three-dimensional laser scanning technology, obtaining three-dimensional point cloud data of the first oil-gas platform as point cloud data to be denoised and geometrical space information of the point cloud data to be denoised, and simultaneously scanning a second oil-gas platform as a sample object by using three-dimensional laser to obtain three-dimensional point cloud data of the second oil-gas platform as point cloud data for a subsequent training model and obtain geometrical space information of the point cloud data for the subsequent training model;
secondly, preprocessing the original data: the preprocessing process mainly comprises the steps of artificially screening and cutting the point cloud data which need to be denoised and the point cloud data used for a subsequent training model before the point cloud data are input into the model according to a statistical method and model characteristics, and converting the original data into data which do not contain abnormal points and conform to the model input format. The method comprises the following steps:
first step, outlier screening: and respectively removing outliers in the point cloud data to be denoised and the point cloud data of the subsequent training model by using the geometric space information of the three-dimensional point cloud data through outlier analysis.
And secondly, normalizing the data format: because the input convolutional neural network needs a standard data format, the point cloud data needing to be denoised is firstly subjected to standardized slicing, and the data can not be exactly divided by the size of the slice, so that the size of a data superposition area existing between two adjacent slices is determined by setting a superposition parameter alpha of the two adjacent slices according to the size of the data needing to be denoised in advance so as to adapt to different point cloud data to be denoised and point cloud data of a subsequent training model; the method comprises the following specific steps: sequentially cutting the point cloud data needing to be denoised and the point cloud data of the subsequent training model into blocks according to the scanning direction to obtain a plurality of standardized point cloud data block matrixes needing to be denoised and point cloud data block matrixes of the subsequent training model, wherein the size of each data block matrix is A, B and C, A represents the number of data contained in the length direction of the data block, B represents the number of data contained in the width direction of the data block, and C represents the number of data contained in the height direction of the data block, such as: the size of each data block square matrix can be selected from commonly used 1024 × 256 data blocks, the specific size can be determined according to the denoising effect, overlapping regions exist in two adjacent point cloud data block square matrices, the size of the overlapping regions is determined by a set overlapping parameter α, and the size of the overlapping regions is (a × α) B × C, for example: may be (1024 × α) 256 × 256. Typically, the alpha value is selected to be 1% -2% of the block width (1024).
Step three, constructing a convolutional neural network model, and specifically comprising the following processes:
step one, setting hyper-parameters of a convolutional neural network model A through a function in sklern to build the convolutional neural network model A, wherein the hyper-parameters comprise: the convolutional neural network model a input layer size is set to the size of the normalized data block square matrix obtained by slicing, i.e., a × B × C, as: 1024 × 256, specifically, for the input data block of the data to be denoised, the input of the convolutional neural network has a C layer in common, such as: 256 layers, each layer corresponding to a × B array data, such as: 1024 × 256 array data. Therefore, the spatial geometrical structure of the data can be reserved, and the training mode of the ordinary convolutional neural network can be met. The convolutional neural network adopts a structure that convolutional layers and pooling layers are alternated, the depths of the convolutional layers and the pooling layers are all set to be 10, each convolutional layer uses 64 convolutional kernels, activating functions of the convolutional layers and output layers are all relu activating functions with the strongest robustness, square error loss is used as a loss function, and a mode of adapting to learning rate is adopted to automatically adjust the learning rate according to the stage of training. The final output layer is a plurality of gaussian noise categories contained in each data block square matrix with the size of a × B × C (for example: 1024 × 256) after the data block square matrix is cut by the convolutional neural network, a characteristic number i represents gaussian noise of different categories, i is 1,2,3 … … … n, and the value of n can be selected to be the most suitable value according to experimental tests, for example: and n is 10, and characteristic numbers 1-10 represent Gaussian noise of different classes respectively.
Secondly, adopting the method of the first step to re-establish n convolutional neural network models B with the same structure as the convolutional neural network model B in the first stepiI is 1 to n, and finally each convolutional neural network model BiEach convolution neural network model B is a denoising model obtained by adding Gaussian noise of the same characteristic number category as the value i to a data block square matrix of an input layer and trainingiThe output result of the output layer is a data block square matrix with the Gaussian noise category corresponding to the characteristic number removed;
such as convolutional neural network model B1Is a pair ofAdding Gaussian noise training corresponding to the characteristic number 1 into a data block square matrix of an input layer to obtain a denoising model; convolutional neural network model B2Adding Gaussian noise corresponding to the characteristic number 2 into a data block square matrix of an input layer to train to obtain a denoising model, and so on;
the following is illustrated with 10 gaussian noises:
if 10 noise types are extracted from the data block square matrix in the first step, the characteristic numbers 1-10 are used for representing Gaussian noises of different types, the 1024 × 256 data block square matrix corresponding to each characteristic number is superposed with the Gaussian noises of the type corresponding to the characteristic number and then input into a convolutional neural network model for training again, and 10 convolutional neural networks with the same structure and different parameters are obtained after training. The method aims to denoise the data blocks with similar noise characteristics by using the same convolutional neural network, and the data blocks with different characteristics can be denoised by using different convolutional neural networks, so that the obtained convolutional neural network is only suitable for samples of corresponding characteristic numbers, and the condition that a certain part of parameter values are excessively adjusted due to individual noise is avoided, so that the denoising effect on other data is not ideal.
And adding the model obtained in the first step, namely 11 convolutional neural network models need to be established, wherein the convolutional neural network model A is used for classifying the noise contained in different data blocks, and the convolutional neural network numbered 1-10 is used for inputting the classified data blocks and completing the denoising of the classified data blocks.
Step four, training a convolutional neural network model, and the specific process is as follows:
the first step, training data preparation:
performing data enhancement on the point cloud data block square matrix of the subsequent training model, and simultaneously independently superposing different levels of Gaussian noise represented by each characteristic number to all data block square matrices obtained after the point cloud data of the subsequent training model is subjected to block cutting standardization;
the principle of the first step is as follows:
most machine learning methods require a huge number of samples to support the accuracy of the model, since the amount of three-dimensional point cloud data obtained by directly scanning with laser is not enough to make the parameters in the convolutional neural network model converge, therefore, the normalized data with the size of 1024 x 256 obtained by slicing the three-dimensional data obtained by scanning the second petroleum platform by the three-dimensional laser scanning technology according to the spatial position in the first step is subjected to data enhancement, that is, all data blocks of 1024 × 256 obtained by standardization are converted by rotation, expansion, symmetry, etc. to obtain several times of original data, and ten kinds of Gaussian noises with different levels are respectively numbered as 1 to 10, for each numbered gaussian noise, it is superimposed separately on all 1024 × 256 data blocks, resulting in ten types of data with different classes of gaussian noise superimposed.
And secondly, training a convolutional neural network model A. And inputting all the data block matrixes obtained in the first step after data enhancement and noise superposition into a convolutional neural network model A, and training the convolutional neural network A by taking the noise category contained in the input data block matrixes as output to obtain the trained convolutional neural network A.
Thirdly, training each convolution neural network model BiThe specific method comprises the following steps:
firstly, respectively superposing certain class of Gaussian noise in all data block matrixes of a subsequent training model, and then respectively inputting the Gaussian noise into a convolution neural network model B with the same i value and the same characteristic number of the class of Gaussian noiseiIn the method, all data block matrixes (data blocks which are considered to be free of noise) of the subsequent training model before the Gaussian noise is superimposed are used as each convolutional neural network model BiTo the convolutional neural network model BiTraining to obtain a trained convolutional neural network Bi。
And step two, continuously repeating the step one until all the convolutional neural network models B are completediTraining (i is 1 to n);
and fifthly, denoising by using the convolutional neural network model, wherein the specific process is as follows:
firstly, inputting all standardized square matrix data of point cloud data to be denoised into a trained convolutional neural network model A, and classifying Gaussian noise contained in the standardized square matrix data of the point cloud data to be denoised by using the convolutional neural network A.
Secondly, respectively inputting the standardized data block matrixes of all point cloud data to be denoised with the same characteristic number of noise category classification into a trained convolutional neural network model B with the same value of i as the characteristic number of Gaussian noise of the categoryiMiddle, convolution neural network model BiOutputting all data block matrixes after removing the noise corresponding to the feature numbers of the noise category classification;
step six, restoring the de-noised data:
and C, performing edge matching on the denoised data block square matrix obtained in the step five, and matching a certain denoised data block square matrix to the data block square matrix adjacent to the certain denoised data block square matrix by utilizing a query algorithm according to the characteristic that the certain denoised data block square matrix is overlapped with the edge data of the adjacent data block square matrix. And splicing the adjacent data block matrixes according to the position of the data of the overlapped part. The specific method comprises the following steps:
the method comprises the steps of firstly, randomly selecting a denoised first data block square matrix;
secondly, matching a second data block square matrix which is overlapped with the edge data of the first data block square matrix by using a query algorithm;
thirdly, splicing the overlapped part of the first data block square matrix and the second data block square matrix;
and fourthly, repeating the second step and the third step, continuously splicing the adjacent data block matrixes of the current data block matrix by using a cyclic algorithm, and finally splicing all the data block matrixes to obtain point cloud data which is denoised and has the same structure as the input original noisy data.
Claims (1)
1. A point cloud data denoising method based on a convolutional neural network is characterized by comprising the following steps:
the method comprises the following steps of firstly, acquiring original three-dimensional point cloud data: scanning a first oil-gas platform needing to obtain a three-dimensional configuration in an actual project by adopting a three-dimensional laser scanning technology, obtaining three-dimensional point cloud data of the first oil-gas platform as point cloud data to be denoised and geometrical space information of the point cloud data to be denoised, and simultaneously scanning a second oil-gas platform as a sample object by using three-dimensional laser to obtain three-dimensional point cloud data of the second oil-gas platform as point cloud data for a subsequent training model and obtain geometrical space information of the point cloud data for the subsequent training model;
step two, preprocessing the original data, comprising the following steps:
first step, outlier screening: respectively removing outliers in the point cloud data to be denoised and the point cloud data of the subsequent training model by using the geometric space information of the three-dimensional point cloud data through outlier analysis;
and step two, standardizing a data format, which comprises the following specific steps: sequentially cutting the point cloud data to be denoised and the point cloud data of a subsequent training model into blocks according to the scanning direction to obtain a plurality of standardized point cloud data block matrixes to be denoised and point cloud data block matrixes of the subsequent training model, wherein the size of each data block matrix is A & ltB & gt & gtC, A represents the number of data contained in the length direction of the data block, B represents the number of data contained in the width direction of the data block, C represents the number of data contained in the height direction of the data block, two adjacent point cloud data block matrixes have a superposition region, the size of the superposition region is determined by a set superposition parameter alpha, and the size of the superposition region is (A & ltalpha & gt & ltalpha & gtB & ltC & gt);
step three, constructing a convolutional neural network model, and specifically comprising the following processes:
step one, setting hyper-parameters of a convolutional neural network model A through a function in sklern to build the convolutional neural network model A, wherein the hyper-parameters comprise: the method comprises the steps that the size of an input layer of a convolutional neural network model A is set to be the size of a standardized data block square matrix obtained through block cutting, namely A, B, C, a convolutional layer and a pooling layer are arranged in the convolutional neural network in an alternating structure, the depth of the convolutional layer and the pooling layer is set to be 10, each convolutional layer uses 64 convolutional kernels, the activation functions of the convolutional layer and an output layer select relu activation functions, square error loss is used as a loss function, a learning rate is automatically adjusted in a mode of adapting to the learning rate according to the training, finally the output layer is various Gaussian noise types contained in the data block square matrix with the size of A, B, C and the size of each data block square matrix extracted by the convolutional neural network and subjected to block cutting, a characteristic number i represents different types of Gaussian noise, and i is 1,2 and 3 … … … n;
secondly, adopting the method of the first step to re-establish n convolutional neural network models B with the same structure as the convolutional neural network model B in the first stepiI is 1 to n, and finally each convolution neural network model BiEach convolution neural network model B is a denoising model obtained by adding Gaussian noise of the same characteristic number category as the value i to a data block square matrix of an input layer and trainingiThe output result of the output layer is a data block square matrix with the Gaussian noise category corresponding to the characteristic number removed;
step four, training a convolutional neural network model, and the specific process is as follows:
the method comprises the following steps of firstly, training data preparation:
performing data enhancement on the point cloud data block square matrix of the subsequent training model, and simultaneously independently superposing different levels of Gaussian noise represented by each characteristic number to all data block square matrices obtained after the point cloud data of the subsequent training model is diced and standardized;
secondly, training a convolutional neural network model A, wherein the specific method comprises the following steps: inputting a convolutional neural network model A by using all the data block matrixes obtained in the first step after data enhancement and noise superposition, and training the convolutional neural network A by taking the noise category contained in the input data block matrixes as output to obtain a trained convolutional neural network A;
thirdly, training each convolution neural network model BiThe specific method comprises the following steps:
firstly, respectively superposing certain class of Gaussian noise in all data block matrixes of a subsequent training model, and then respectively inputting the Gaussian noise into a convolution neural network model B with the same i value and the same characteristic number of the class of Gaussian noiseiIn the method, the post-training model before Gaussian noise is superimposedWith a square matrix of data blocks as each convolutional neural network model BiFor the convolutional neural network model BiTraining to obtain a trained convolutional neural network Bi；
And step two, continuously repeating the step one until all the convolutional neural network models B are completediTraining;
and fifthly, denoising by using the convolutional neural network model, wherein the specific process is as follows:
firstly, inputting all standardized square matrix data of point cloud data to be denoised into a trained convolutional neural network model A respectively, and classifying Gaussian noise contained in the standardized square matrix data of the point cloud data to be denoised by using the convolutional neural network A;
secondly, respectively inputting the standardized data block square matrix of all point cloud data to be denoised with the same characteristic number of noise category classification into a trained convolutional neural network model B with the same i value as the characteristic number of Gaussian noise of the categoryiMiddle, convolution neural network model BiOutputting all data block matrixes after removing noise corresponding to the feature numbers of the noise category classification;
and step six, restoring the de-noised data, wherein the specific method comprises the following steps:
the method comprises the steps of firstly, randomly selecting a denoised first data block square matrix;
secondly, matching a second data block square matrix which is overlapped with the edge data of the first data block square matrix by using a query algorithm;
thirdly, splicing the data overlapped part of the first data block square matrix and the second data block square matrix;
and fourthly, repeating the second step and the third step, continuously splicing the adjacent data block matrixes of the current data block matrix by using a cyclic algorithm, and finally splicing all the data block matrixes to obtain point cloud data which is denoised and has the same structure as the input original noisy data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011538245.1A CN112669230B (en) | 2020-12-23 | 2020-12-23 | Point cloud data denoising method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011538245.1A CN112669230B (en) | 2020-12-23 | 2020-12-23 | Point cloud data denoising method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112669230A CN112669230A (en) | 2021-04-16 |
| CN112669230B true CN112669230B (en) | 2022-07-08 |
Family
ID=75408490
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011538245.1A Active CN112669230B (en) | 2020-12-23 | 2020-12-23 | Point cloud data denoising method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112669230B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117337449A (en) * | 2021-04-28 | 2024-01-02 | Oppo广东移动通信有限公司 | Point cloud quality enhancement method, encoding and decoding methods and devices, and storage medium |
| CN113191973B (en) * | 2021-04-29 | 2023-09-01 | 西北大学 | Cultural relic point cloud data denoising method based on unsupervised network framework |
| CN116188719B (en) * | 2023-04-27 | 2023-11-17 | 中地云智慧科技有限公司 | Fragment removing method for live-action three-dimensional data |
| CN118015284B (en) * | 2024-04-09 | 2024-06-25 | 中国空气动力研究与发展中心低速空气动力研究所 | Point cloud segmentation method and system based on generation countermeasure network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110136072A (en) * | 2018-02-08 | 2019-08-16 | 北京大学 | Minimizing technology, denoising system, computer equipment and the storage medium of point cloud noise |
| CN111640072A (en) * | 2020-05-09 | 2020-09-08 | 中国地质大学（武汉） | Point cloud denoising method combining feature detection method and vertex updating method |
| CN112101278A (en) * | 2020-09-25 | 2020-12-18 | 湖南盛鼎科技发展有限责任公司 | Hotel point cloud classification method based on k nearest neighbor feature extraction and deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019133922A1 (en) * | 2017-12-29 | 2019-07-04 | Flir Systems, Inc. | Point cloud denoising systems and methods |
| CN109949235A (en) * | 2019-02-26 | 2019-06-28 | 浙江工业大学 | A kind of chest x-ray piece denoising method based on depth convolutional neural networks |
-
2020
- 2020-12-23 CN CN202011538245.1A patent/CN112669230B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110136072A (en) * | 2018-02-08 | 2019-08-16 | 北京大学 | Minimizing technology, denoising system, computer equipment and the storage medium of point cloud noise |
| CN111640072A (en) * | 2020-05-09 | 2020-09-08 | 中国地质大学（武汉） | Point cloud denoising method combining feature detection method and vertex updating method |
| CN112101278A (en) * | 2020-09-25 | 2020-12-18 | 湖南盛鼎科技发展有限责任公司 | Hotel point cloud classification method based on k nearest neighbor feature extraction and deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112669230A (en) | 2021-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112669230B (en) | Point cloud data denoising method based on convolutional neural network | |
| CN110555474B (en) | Photovoltaic panel fault detection method based on semi-supervised learning | |
| CN110705457B (en) | Remote sensing image building change detection method | |
| CN110532859B (en) | Remote sensing image target detection method based on deep evolution pruning convolution net | |
| CN111815601B (en) | Texture image surface defect detection method based on depth convolution self-encoder | |
| CN112837295B (en) | Rubber glove defect detection method based on generation countermeasure network | |
| CN112200244B (en) | Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training | |
| CN109858408B (en) | Ultrasonic signal processing method based on self-encoder | |
| CN112560918B (en) | Dish identification method based on improved YOLO v3 | |
| CN109410149B (en) | CNN denoising method based on parallel feature extraction | |
| CN111815041B (en) | Seawater temperature field prediction method based on improved EOF | |
| CN115115921B (en) | Wafer defect mode identification method based on residual network | |
| CN113408423A (en) | Aquatic product target real-time detection method suitable for TX2 embedded platform | |
| CN110568483A (en) | Automatic evaluation method for seismic linear noise suppression effect based on convolutional neural network | |
| CN112232229A (en) | Fine water body extraction method based on U-net neural network | |
| CN114283120B (en) | Domain-adaptive-based end-to-end multisource heterogeneous remote sensing image change detection method | |
| CN116229380A (en) | Method for identifying bird species related to bird-related faults of transformer substation | |
| CN110297401A (en) | Extreme ultraviolet photolithographic mask multilayer film phase type defect inspection method | |
| CN110334478A (en) | Machinery equipment abnormality detection model building method, detection method and model | |
| CN115223054A (en) | Remote sensing image change detection method based on partition clustering and convolution | |
| CN111488974B (en) | Ocean wind energy downscaling method based on deep learning neural network | |
| CN112421631A (en) | New energy consumption capacity assessment method and system | |
| CN113487600A (en) | Characteristic enhancement scale self-adaptive sensing ship detection method | |
| CN115601572A (en) | Ultrasonic phased array image optimization reconstruction method and system based on semi-supervised CycleGan network | |
| CN116563250A (en) | Recovery type self-supervision defect detection method, device and storage medium |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |