River sluice point cloud denoising method of deep neural network based on reinforcement learning decision

Abstract

The invention discloses a river sluice point cloud denoising method of a deep neural network based on reinforcement learning decision. The method comprises the following steps: 1) constructing a multi-scale feature extraction network; 2) based on the multi-scale feature extraction network constructed in the step 1), constructing a point cloud denoising framework based on reinforcement learning decision, and selecting an optimal denoising path for each point in the point cloud by using the reinforcement learning decision instead of traditional loss self-adaption; and 3) verifying the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision by taking the maximum and minimum distance square MDS and the average maximum and minimum distance square MMDS between the corresponding points of the ground real point cloud and the denoised point cloud as evaluation indexes. According to the method, a training strategy is dynamically adjusted by adopting reinforcement learning, so that the model can adapt to noise modes under different local structures, more accurate noise point displacement estimation is realized, and high-quality point cloud denoising is realized while key details are reserved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of river barrage point cloud denoising, in particular to a river barrage point cloud denoising method based on deep neural network of reinforcement learning decision. BACKGROUND

[0002] In the process of three-dimensional laser scanning of the surface of the river barrage, due to environmental interference, equipment precision limitation and human operation and other factors, the obtained point cloud data is inevitably contaminated by multiple noises, including random noise (caused by laser scattering and electronic noise), system noise (originated from the inherent precision limitation of the equipment) and occlusion noise (caused by the absence of point cloud due to object occlusion) and the like. These noises not only cause outliers to be mixed therein, but also cause uneven distribution of point cloud density, which seriously affects the subsequent segmentation, registration and three-dimensional reconstruction accuracy, and therefore the point cloud denoising work is the basis for other point cloud post-processing operations.

[0003] The river barrage structure is complex and composed of multiple components, and the surface curvature of each component changes significantly, resulting in obvious unevenness of the scanned point cloud density. Traditional point cloud denoising algorithms (such as radius filtering, statistical filtering and median filtering) highly depend on parameters such as neighborhood selection, maximum distance threshold and acceptance error, and these parameters are difficult to be uniformly set under the complex surface structure of the river barrage, often leading to unstable denoising effect, and even causing data over-smoothing or loss of key features. Although the point cloud denoising method based on deep learning has made significant progress in recent years, the existing methods mostly use a simple measurement based on the nearest distance between noisy points and noise-free points as the training target, ignoring the local structural features of the point cloud, resulting in inaccurate displacement estimation, and the denoised point cloud often has problems such as abnormal values, shrinkage and aggregation. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the present application provides a river barrage point cloud denoising method based on deep neural network of reinforcement learning decision, which optimizes the training strategy by introducing a reinforcement learning mechanism to achieve good denoising effect. The method discards the idea of global uniform processing and focuses on local area feature extraction. By adaptively learning the local geometric characteristics of the point cloud, the denoising ability for complex surface structures of the river barrage is significantly improved, and the local geometric features of the point cloud can be effectively captured. At the same time, the training strategy is dynamically adjusted by using reinforcement learning, so that the model can adapt to different noise patterns under different local structures, realize more accurate noise point displacement estimation, and thus realize high-quality point cloud denoising while preserving key details.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows.

[0006] A river barrage point cloud denoising method based on deep neural network of reinforcement learning decision, comprising the following steps:

[0007] Step S1, constructing a multi-scale feature extraction network;

[0008] The multi-scale feature extraction network comprises a spatial transformation network and a feedforward neural network. The normalized point cloud data is input into the network, first passes through the spatial transformation network to obtain transformed features, then the transformed features are fused with the normalized features, and then input into the feedforward neural network for dimensionality expansion. After repeating the above process once, the obtained dimensionality expanded features are summed and input into the feedforward neural network to obtain the final result.

[0009] Step S2, constructing a point cloud denoising framework based on reinforcement learning decision based on the multi-scale feature extraction network constructed in step S1;

[0010] The point cloud denoising framework based on reinforcement learning decision comprises an encoder, a point cloud denoising block, a decoder and a path decision intelligent agent. Reinforcement learning decision is used to replace the traditional loss to adaptively select the best denoising path for each point in the point cloud.

[0011] Step S3, taking the maximum distance square MDS and the average maximum distance square MMDS between the ground true point cloud and the corresponding points of the denoised point cloud as evaluation indexes, verifying the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision.

[0012] Specifically, the process of constructing the multi-scale feature extraction network in step S1 is as follows:

[0013] Step S11, constructing a spatial transformation network suitable for point cloud data, which comprises at least a positioning grid and a grid generator;

[0014] The positioning grid receives an input feature map with a batch size B, a channel C and a length L and outputs , that is, the parameters of the transformation applied to the feature map: =floc( ); The size of changes according to the parameterized transformation type:

[0015] (1)

[0016] In the above formula, , , is the original information, , , is the transformed information, is the transformation parameter;

[0017] The affine transformation corresponding to the tensor of the three dimensions in formula (1) in turn​ The affine transformation corresponding to a 2-dimensional tensor of size [3, 4]. The size is [2, 3], and so on, with corresponding changes.

[0018] floc() takes a fully connected or convolutional network and includes at least one final regression layer to generate the transformation parameters. ;

[0019] pass The parameter group can perform rotation, scaling, and translation operations on a given input feature map. After transformation , The transformation parameters generated by the localization network are used to output features. When the output feature is defined as located in the feature Regular grid When the above is done, an output feature map is formed. This leads to a coordinate matrix transformation relationship:

[0020] (2)

[0021] (3)

[0022] In the above formula, and These are the coordinates of the output target point cloud. and These are the coordinates of the origin cloud. Indicates an affine relationship;

[0023] The mapping relationship of the coordinate matrix is ​​from the target feature to the input feature. The purpose of the grid generator is to sample the target feature on the original feature. Each time, pixels are sampled from different positions of the original feature and transferred to the target feature, thus covering the target feature. The position of the target feature is fixed in each sampling process, while the position of the sampled original feature is not fixed.

[0024] Step S12: Construct a feedforward neural network. The feedforward neural network includes at least one input layer, one output layer, and multiple hidden layers. The neurons in the previous layer and the neurons in the next layer of the feedforward neural network are all fully connected.

[0025] The complete training cycle of the feedforward neural network includes the following process:

[0026] Initialization parameters: The weights w and biases b of the feedforward neural network are randomly initialized during the first iteration;

[0027] Forward propagation: During training, the input data is propagated forward through multiple fully connected layers. The input data is linearly combined with the weights and biases of each layer, and then a non-linear activation function is used to improve the overall expressive power.

[0028] Backpropagation: The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters. Starting from the output layer, the partial derivatives of the loss with respect to the parameters are calculated layer by layer, and this gradient information is propagated from the output layer back to the input layer.

[0029] Step S13: Construct a residual neural network;

[0030] The residual neural network is used to address the problem of decreased network expressive power, ensuring that the expressive power of a more layered network model is not lower than that of a less layered network model. The expression of the residual neural network is as follows:

[0031] (4)

[0032] In the above formula, This is the output vector; The input vector; This represents the residual mapping.

[0033] Furthermore, the specific process of backpropagation in step S12 is as follows:

[0034] Step S121: Calculate the loss;

[0035] The error between the predicted output and the true label is calculated using a loss function. The mathematical expression for the loss function is:

[0036] (5)

[0037] In the above formula, n is the number of samples; This is the actual value; This is a predicted value;

[0038] Considering that the noisy input data may contain outliers in the actual calculation process, the loss function given by formula (5) is greatly affected by outliers, which may lead to the inability to find a suitable gradient. Therefore, the average distance between the model's predicted value and the sample's true value is used as the loss function, and its expression is:

[0039] (6)

[0040] Step S122: Gradient calculation;

[0041] The gradient of the loss function with respect to each weight and bias is calculated using the chain rule;

[0042] Step S123: Weight update;

[0043] Each weight and bias is updated using the gradient descent algorithm to reduce error. The calculation formula is as follows:

[0044] (7)

[0045] In the above formula, It is weight; L is the learning rate; L is the loss function. It is the partial derivative of the loss function with respect to the weights;

[0046] Step S124: Parameter update;

[0047] After obtaining the gradient, the Adam optimizer is used to update the model parameters. Adam uses a window decay accumulation of first-order momentum, expressed by the following formula:

[0048] (8)

[0049] In the above formula, The first-order momentum of the current step; This is the first-order momentum of the previous step; The historical first-order momentum retention rate;

[0050] Adam uses a window decay accumulation of second-order momentum, expressed by the following formula:

[0051] (9)

[0052] In the above formula, Let be the second momentum of the current step; This is the second momentum of the previous step; The decay rate of the historical second-order momentum;

[0053] The final update formula is:

[0054] (10)

[0055] In the above formula, This represents the amount of weight update; The learning rate; The first-order momentum of the current step; Let be the second momentum of the current step; A coefficient to increase the stability of the denominator.

[0056] Specifically, step S2 involves constructing a point cloud denoising framework based on reinforcement learning decision-making, as detailed below:

[0057] Set a point cloud set N, For point cloud collection The point cloud set after adding noise is used for point cloud denoising, which involves finding the set. and set The 3D geometric relationship of the midpoint cloud, noise set The expression is as follows:

[0058] (11)

[0059] In the above formula, It is a set of Gaussian noise; It is a collection of point clouds containing noise;

[0060] Traditional noise reduction methods are effective against noise. Typically, indiscriminate removal is performed, while noise... The form of point cloud data is usually uncertain, and the three-dimensional shape and scale of point cloud sets are also diverse. Using a single neural network model to denoise point cloud data will result in noise residue, over-smoothing, and shape distortion. To address this, this invention designs a reinforcement learning-based decision-making method to adaptively select the best denoising path for each point in the point cloud.

[0061] The encoder will output noisy point clouds. Normalized to a unit sphere centered at the origin, for each point in the point cloud. Construct local regions and set different sphere query radii or use multiple layers with different resolutions to extract features;

[0062] The decoder is positioned after the last denoising block, and the decoder processes the shift vector. Regression calculations are performed, and the point cloud is output after max pooling and a fully connected layer to remove noise.

[0063] The point cloud denoising block includes multiple stacked denoising blocks. The denoising block is first connected to a shared multilayer perceptron, and then merged with the original features before entering the agent decision-making process. The denoising block contains three denoising paths: the first is to directly output the features; the second is a multilayer perceptron MLP; and the third is a multi-scale feature extraction network. An agent based on reinforcement learning decision-making selects a path for denoising.

[0064] The agent based on reinforcement learning decision-making includes a multilayer perceptron (MLP), a max pooling layer, a long short-term memory (LSTM) module, a feedforward neural network (FNN), and activation functions. Traditional deep learning algorithms determine the training direction through loss, which cannot achieve ideal results in path selection. Therefore, reinforcement learning decision-making is used to replace traditional loss.

[0065] Suppose the state of the given reinforcement network is Network actions Network intelligent agents are The relationship between the three is as follows:

[0066] (12)

[0067] In the above formula, For intelligent agents; To provide status based on the network environment The action made at that time;

[0068] Based on the above, further network rewards will be introduced. The initial noise level is incorporated into the reward to guide the agent to pay more attention to noise information; network reward The formula is expressed as:

[0069] (13)

[0070] In the above formula, This is the action in reinforcement learning, and its value is either 0 or 1. Penalty parameters for path selection; The loss between the original noise and the clean data; The loss between the denoised results and the clean data; Noise level; For scaling parameters; These are the basic weight parameters.

[0071] Specifically, in step S3, the squared maximum-minimum distance (MDS) and squared average maximum-minimum distance (MMDS) between corresponding points in the ground truth point cloud and the denoised point cloud are expressed by the following formulas:

[0072] (14)

[0073] (15)

[0074] In the above formula, It is a real point cloud; It is a point cloud after noise reduction; These are points in a real point cloud; These are the points in the point cloud after denoising. Represents the corresponding point cloud after denoising of Nearest neighbor points, Set to 10; This represents the actual number of point clouds on the ground; both MDS and MMDS are measured in mm.

[0075] The smaller the values ​​of MDS and MMDS, the better the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision-making.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] 1. The method of this invention abandons the idea of ​​global unified processing and focuses on local feature extraction. By adaptively learning the local geometric characteristics of point clouds, it significantly improves the noise reduction capability of complex surface structures of river dams and can effectively capture the local geometric features of point clouds.

[0078] 2. The method of this invention adopts a reinforcement learning dynamic adjustment training strategy, which enables the model to adapt to noise patterns under different local structures, achieve more accurate noise point displacement estimation, and thus achieve high-quality point cloud denoising while preserving key details. Attached Figure Description

[0079] Figure 1 This is a flowchart of the point cloud denoising method for river dams based on deep neural networks for reinforcement learning decision-making, as described in this invention.

[0080] Figure 2 This is a schematic diagram of the spatial transformation network structure of the present invention;

[0081] Figure 3 This is a schematic diagram of the feedforward neural network structure of the present invention;

[0082] Figure 4 This is a schematic diagram of the residual module structure of the present invention;

[0083] Figure 5 This is a schematic diagram of the complete network structure of the multi-scale feature extraction network of the present invention;

[0084] Figure 6 This is a schematic diagram of the network structure for reinforcement learning decision-making based on the present invention;

[0085] Figure 7 This is a schematic diagram of the network structure of the noise reduction block of the present invention;

[0086] Figure 8 This is a schematic diagram of the path decision-making agent network structure of the present invention;

[0087] Figure 9 This is a schematic diagram of the river dam site in an embodiment of the present invention;

[0088] Figure 10 This is a diagram showing the on-site layout of the river dam data acquisition operation in an embodiment of the present invention;

[0089] Figure 11 This is a schematic diagram showing the distribution of the monitoring stations at the river dam in this embodiment of the invention;

[0090] Figure 12 This is a comparison of the visualization results before and after denoising in the dam working bridge-1 scenario under 2% noise in an embodiment of the present invention;

[0091] Figure 13 This is a comparison of the visualization results before and after denoising in the 2% noise level scene of the working bridge-2 of the river dam in this embodiment of the invention;

[0092] Figure 14 This is a comparison of the visualization results before and after denoising in the dam traffic bridge-1 scene under 2% noise in an embodiment of the present invention;

[0093] Figure 15 This is a comparison of the visualization results before and after denoising in the 2% noise level scene of the river dam traffic bridge-2 in this embodiment of the invention;

[0094] Figure 16 This is a comparison of the visualization results before and after noise reduction of the dam pier scene under 2% noise in an embodiment of the present invention. Detailed Implementation

[0095] To facilitate understanding and implementation of the present invention by those skilled in the art, the various steps of the method proposed in this invention are described in detail below. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0096] Example 1

[0097] like Figure 1 As shown, this invention discloses a method for denoising point clouds of river dams based on deep neural networks using reinforcement learning decision-making, comprising the following steps:

[0098] Step S1: Construct a multi-scale feature extraction network;

[0099] like Figure 5 As shown, the multi-scale feature extraction network includes a spatial transformation network and a feedforward neural network. Normalized point cloud data is input into this network. First, it passes through the spatial transformation network to obtain transformed features. Then, the transformed features are fused with the normalized features and subsequently input into the feedforward neural network to increase their dimensionality. After repeating the above process once, the resulting increased-dimensional features are summed and input into the feedforward neural network to obtain the final result.

[0100] Step S2: Based on the multi-scale feature extraction network constructed in step S1, construct a point cloud denoising framework based on reinforcement learning decision-making.

[0101] The point cloud denoising framework based on reinforcement learning decision-making includes an encoder, a point cloud denoising block, a decoder, and a path decision agent. It uses reinforcement learning decision-making to replace traditional loss and adaptively selects the best denoising path for each point in the point cloud.

[0102] Step S3: Using the squared extreme distance between the ground real point cloud and the corresponding points in the denoised point cloud as evaluation metrics, verify the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision-making.

[0103] Specifically, the process of constructing the multi-scale feature extraction network in step S1 is as follows:

[0104] Step S11: Construct a spatial transformation network suitable for point cloud data, such as... Figure 2 As shown, a spatial transformation network includes at least a positioning grid and a grid generator;

[0105] The positioning grid receives input feature maps with a batch size of B, a channel of C, and a length of L. and output That is, the transformation applied to the feature map Parameters: =floc( ); The size varies accordingly based on the type of parameterized transformation:

[0106] (1)

[0107] In the above formula, , , This is the original information. , , For the transformed information, For transformation parameters;

[0108] In formula (1), the affine transformations corresponding to the tensors of the three dimensions are given in sequence. The affine transformation corresponding to a 2-dimensional tensor of size [3, 4]. The size is [2, 3], and so on, with corresponding changes.

[0109] floc() takes a fully connected or convolutional network and includes at least one final regression layer to generate the transformation parameters. ;

[0110] pass The parameter group can perform rotation, scaling, and translation operations on a given input feature map. After transformation , The transformation parameters generated by the localization network are used to output features. When the output feature is defined as located in the feature Regular grid When the above is done, an output feature map is formed. This leads to a coordinate matrix transformation relationship:

[0111] (2)

[0112] (3)

[0113] In the above formula, and These are the coordinates of the output target point cloud. and These are the coordinates of the origin cloud. Indicates an affine relationship;

[0114] The mapping relationship of the coordinate matrix is ​​from the target feature to the input feature. The purpose of the grid generator is to sample the target feature on the original feature. Each time, pixels are sampled from different positions of the original feature and transferred to the target feature, thus covering the target feature. The position of the target feature is fixed in each sampling process, while the position of the sampled original feature is not fixed.

[0115] Step S12: Construct a feedforward neural network, such as... Figure 3 As shown, a feedforward neural network includes at least one input layer, one output layer, and multiple hidden layers. The neurons in the previous layer and the neurons in the next layer of the feedforward neural network are all fully connected.

[0116] The complete training cycle of the feedforward neural network includes the following process:

[0117] Initialization parameters: The weights w and biases b of the feedforward neural network are randomly initialized during the first iteration;

[0118] Forward propagation: During training, the input data is propagated forward through multiple fully connected layers. The input data is linearly combined with the weights and biases of each layer, and then a non-linear activation function is used to improve the overall expressive power.

[0119] Backpropagation: The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters. Starting from the output layer, the partial derivatives of the loss with respect to the parameters are calculated layer by layer, and this gradient information is propagated from the output layer back to the input layer.

[0120] Step S13: Construct a residual neural network;

[0121] like Figure 4 As shown, the residual neural network is used to address the problem of decreased network expressive power, ensuring that the expressive power of a more layered network model is not lower than that of a less layered network model. The expression of the residual neural network is as follows:

[0122] (4)

[0123] In the above formula, This is the output vector; The input vector; This represents the residual mapping.

[0124] Furthermore, the specific process of backpropagation in step S12 is as follows:

[0125] Step S121: Calculate the loss;

[0126] The error between the predicted output and the true label is calculated using a loss function. The mathematical expression for the loss function is:

[0127] (5)

[0128] In the above formula, n is the number of samples; This is the actual value; This is a predicted value;

[0129] Considering that the noisy input data may contain outliers in the actual calculation process, the loss function given by formula (5) is greatly affected by outliers, which may lead to the inability to find a suitable gradient. Therefore, the average distance between the model's predicted value and the sample's true value is used as the loss function, and its expression is:

[0130] (6)

[0131] Step S122: Gradient calculation;

[0132] The gradient of the loss function with respect to each weight and bias is calculated using the chain rule;

[0133] Step S123: Weight update;

[0134] Each weight and bias is updated using the gradient descent algorithm to reduce error. The calculation formula is as follows:

[0135] (7)

[0136] In the above formula, It is weight; L is the learning rate; L is the loss function. It is the partial derivative of the loss function with respect to the weights;

[0137] Step S124: Parameter update;

[0138] After obtaining the gradient, the Adam optimizer is used to update the model parameters. Adam uses a window decay accumulation of first-order momentum, expressed by the following formula:

[0139] (8)

[0140] In the above formula, The first-order momentum of the current step; This is the first-order momentum of the previous step; The historical first-order momentum retention rate;

[0141] Adam uses a window decay accumulation of second-order momentum, expressed by the following formula:

[0142] (9)

[0143] In the above formula, Let be the second momentum of the current step; This is the second momentum of the previous step; The decay rate of the historical second-order momentum;

[0144] The final update formula is:

[0145] (10)

[0146] In the above formula, This represents the amount of weight update; The learning rate; The first-order momentum of the current step; Let be the second momentum of the current step; A coefficient to increase the stability of the denominator.

[0147] Specifically, step S2 involves constructing a point cloud denoising framework based on reinforcement learning decision-making, as detailed below:

[0148] Set a point cloud set N, For point cloud collection The point cloud set after adding noise is used for point cloud denoising, which involves finding the set. and set The 3D geometric relationship of the midpoint cloud, noise set The expression is as follows:

[0149] (11)

[0150] In the above formula, It is a set of Gaussian noise; It is a collection of point clouds containing noise;

[0151] Traditional noise reduction methods are effective against noise. Typically, indiscriminate removal is performed, while noise... The form of point cloud data is often uncertain, and the 3D shape and scale of point cloud sets are also diverse. Using a single neural network model to denoise point cloud data can lead to problems such as residual noise, over-smoothing, and shape distortion. Figure 6 As shown, this invention designs a reinforcement learning-based decision-making method to adaptively select the best denoising path for each point in a point cloud;

[0152] The encoder will output noisy point clouds. Normalized to a unit sphere centered at the origin, for each point in the point cloud. Construct local regions and set different sphere query radii or use multiple layers with different resolutions to extract features;

[0153] The decoder is positioned after the last denoising block, and the decoder processes the shift vector. Regression calculations are performed, and the point cloud is output after max pooling and a fully connected layer to remove noise.

[0154] like Figure 7 As shown, the point cloud denoising block includes multiple stacked denoising blocks. The denoising block is first connected to a shared multilayer perceptron, and then merged with the original features before entering the agent decision-making process. The denoising block contains three denoising paths: the first is to directly output the features; the second is a multilayer perceptron MLP; and the third is a multi-scale feature extraction network. A path is selected by an agent based on reinforcement learning decision-making for denoising.

[0155] like Figure 8 As shown, the agent based on reinforcement learning decision-making includes a multilayer perceptron (MLP), a max pooling layer, a long short-term memory (LSTM) module, a feedforward neural network (FNN), and activation functions. Traditional deep learning algorithms determine the training direction through loss, which cannot achieve ideal results in path selection. Therefore, reinforcement learning decision-making is used to replace traditional loss.

[0156] Suppose the state of the given reinforcement network is Network actions Network intelligent agents are The relationship between the three is as follows:

[0157] (12)

[0158] In the above formula, For intelligent agents; To provide status based on the network environment The action made at that time;

[0159] Based on the above, further network rewards will be introduced. The initial noise level is incorporated into the reward to guide the agent to pay more attention to noise information; network reward The formula is expressed as:

[0160] (13)

[0161] In the above formula, This is the action in reinforcement learning, and its value is either 0 or 1. Penalty parameters for path selection; This represents the loss between the original noise and the clean data; The loss between the denoised results and the clean data; Noise level; For scaling parameters; These are the basic weight parameters.

[0162] Specifically, in step S3, the squared maximum-minimum distance (MDS) and squared average maximum-minimum distance (MMDS) between corresponding points in the ground truth point cloud and the denoised point cloud are expressed by the following formulas:

[0163] (14)

[0164] (15)

[0165] In the above formula, It is a real point cloud; It is a point cloud after noise reduction; These are points in a real point cloud; These are the points in the point cloud after denoising. Represents the corresponding point cloud after denoising of Nearest neighbor points, Set to 10; This represents the actual number of point clouds on the ground; both MDS and MMDS are measured in mm.

[0166] The smaller the values ​​of MDS and MMDS, the better the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision-making.

[0167] The technical effects of the method of the present invention will be further illustrated below through a point cloud denoising example of a prototype dam project.

[0168] Brief introduction to the prototype project of the river dam:

[0169] The prototype engineering data for this example is a water conservancy project in Jiangxi Province, which was initially constructed in 1958. This project consists of a dam, intake gate, and ship lock. This example collects 3D point cloud data of the dam, which has 7 spans, each with a net width of 12m. Figure 9 These are photos taken at the site of the river dam. Figure 9(a) shows the actual view upstream of the dam. Figure 9 Image (b) shows the downstream view of the dam. This invention uses a Leica P50 3D laser scanner to acquire point cloud data of the dam and surrounding terrain through multi-station scanning. Figure 10 (a), (b), and (c) in the image show the work site during the data collection for the river dam.

[0170] When scanning the dam, firstly, because the dam was releasing water during the scanning process, the downstream measuring points were placed at a considerable distance from the dam to ensure instrument safety. Secondly, due to the large clear width of each span of the dam and the presence of a weir on the right bank, it was impossible to set up the instrument there. Therefore, the instrument could only be set up on the left bank of the dam to achieve full coverage scanning. However, this resulted in some sections of the dam located far from the measuring station not being scanned at a distance of 120m. For these sections, the scanning parameters were adjusted in this example, and a finer scan was performed at a distance of 270m. For the remaining sections of the dam, a distance of 120m was still selected for scanning to improve data acquisition efficiency. To meet the accuracy requirements of the data required by this invention, a total of 9 stations were set up for this scan: the single-span net width of the dam is relatively large, so 5 stations were set up at the top; the downstream of the left bank of the dam has a wide field of view, so only 1 station was set up; due to the large number of buildings blocking the view upstream of the left bank of the dam and the poor scanning angle, 3 stations were set up upstream of the left bank. Figure 11 This is a schematic diagram showing the distribution of monitoring stations at the dam during data collection.

[0171] I. Experimental Environment of the Invention

[0172] This experiment was based on the PyTorch framework, and the server environment was equipped with an Intel(R) Core™ i7 12700kf CPU, 128 GB RAM, NVIDIA RTX 4090, CUDA 11.7, and cuDNN v8.5.0. The neural network model used the Adam optimizer with an initial learning rate of 0.0001.

[0173] II. Comparison of Noise Reduction Results of Prototype Barrier Gate Components Using Multiple Methods

[0174] In this example, the overall point cloud data of the dam was systematically denoised, and three key components with typical structural features (working bridge, traffic bridge, and dam pier) were selected for visualization and quantitative analysis.

[0175] Figure 12 To compare the visualization results before and after denoising in the dam working bridge-1 scene under 2% noise, Figure 12In the middle (a), it is the Poisson reconstruction image, and (b) is the point cloud image. The subscripts 1-8 are the pure point cloud data-1, the noisy point cloud data-2, the median filtering result-3, the normal bilateral filtering result-4, the mean filtering result-5, the voxel filtering result-6, the PCPNet result-7, and the filtering result of the method of this invention-8, respectively. In (b), the color of the point cloud reflects the magnitude of the MDS value, with small values ​​being blue and large values ​​being red.

[0176] Depend on Figure 11 It can be seen that the working bridge-1 of the dam is composed of 3 archways, 2 pillars and eaves. The Poisson reconstructed image can reflect the merits of each denoising algorithm from these parts. Median filtering (a3) ​​and mean filtering (a5) over-smoothed the various parts of the working bridge-1, resulting in severe distortion of the Poisson reconstructed image. Normal bilateral filtering (a4) and voxel filtering (a6) effectively removed the noise around the archways and pillars, but failed to effectively remove the noise from the area above the archways to the eaves plane. Visually, both the intelligent algorithm PCPNet (a7) and the method of this invention (a8) can better preserve the original information of each surface.

[0177] The larger the MDS value in a point cloud image, the closer the point cloud color is to red. The clean point cloud (b1) and the noisy point cloud (b2) are blue because they have not been denoised. The median filtering method (b3) smooths the geometric features of each part of the scene, and the overall MDS value is large, with only a small number of points being blue. The mean filtering (b5) preserves the geometric features of the scene, but the MDS value of most point clouds is also relatively large. The MDS values ​​obtained by the normal bilateral filtering (b4) and voxel filtering (b6) are generally better than the two classic algorithms, but there are many outliers at the edges. The PCPNet method (b7) and the method of this invention (b8) have similar denoising effects, but the overall MDS value of the method of this invention is smaller than that of the PCPNet method.

[0178] Table 1 below shows the MMDS obtained after denoising using six methods under Gaussian noise with standard deviations of 0.5%, 1%, and 2% of the diagonal length of the point cloud bounding box in the scene of the working bridge-1 of the dam.

[0179] Table 1. MMDS results of noise reduction for the working bridge-1 scene of the river dam.

[0180] ;

[0181] As shown in Table 1, in the scenario of the working bridge-1 of the dam, the method of the present invention outperforms other methods under all noise conditions. When the Gaussian noise has a standard deviation of 0.5%, the MMDS of the method of the present invention is much lower than that of other methods. For example, the MMDS of the median filter is 75 times that of the method of the present invention. The performance of the mean filter is also poor. The performance of the normal bilateral filter is acceptable under low noise conditions, but it is still about 140% higher than that of the method of the present invention. This shows that the accuracy advantage of the method of the present invention in low noise environment is extremely significant.

[0182] As the noise level increases to 1% and 2%, the normal bilateral filtering and voxel filtering methods perform relatively well under low-noise conditions, but the MMDS gradually increases under high-noise conditions, with a significant increase, indicating that they are more suitable for handling low-level noise. The MMDS of median filtering and mean filtering remains consistently high across the entire noise range, indicating that they are not sensitive to changes in noise level. Looking at Table 1 as a whole, the increase in MMDS for the intelligent algorithm PCPNet shows a trend of increasing from small to large: from 0.5% to 1%, the increase is 0.02032 mm; from 1% to 2%, the increase is 2.47 times the previous increase. This indicates that the method is not very robust. The MMDS increase of the method in this invention is relatively small. Overall, traditional methods are prone to failure in high-noise environments, while the method of this invention can effectively cope with complex noise.

[0183] like Figure 13 The image shown is a comparison of the visualization results before and after denoising in the dam working bridge-2 scene under 2% noise. Figure 13 (a) is the Poisson reconstruction image, and (b) is the point cloud image. Subscripts 1-8 represent the following results: clean point cloud data - 1, noisy point cloud data - 2, median filtering result - 3, bilateral normal filtering result - 4, mean filtering result - 5, voxel filtering result - 6, PCPNet result - 7, and the filtering result of the method described in this invention - 8. In (b), the color of the point cloud reflects the magnitude of the MDS value; smaller values ​​are blue, and larger values ​​are red.

[0184] Depend on Figure 12 As can be seen, the working bridge-2 of the river dam is H-shaped, with two main columns and a crossbeam connecting the top and middle. The Poisson reconstructed image reflects the performance of various denoising algorithms through these components. Median filtering (a3) ​​and mean filtering (a5) over-smoothed the working bridge-2, resulting in severe distortion of the Poisson reconstructed image. Bilateral normal filtering (a4) effectively removed the noise point cloud on the scene surface, but over-smoothed the corner point cloud at the transition of the crossbeam. The denoising effects of voxel filtering (a6) and the intelligent algorithm PCPNet (a7) are visually similar, and both can preserve the point cloud information at the corners. The method of this invention (a8) achieves good removal of the overall noise point cloud.

[0185] The larger the MDS value in the point cloud image, the closer the point cloud color is to red. The clean point cloud (b1) and the noisy point cloud (b2) are not denoised, so the point clouds are blue. The median filtering method (b3) excessively smooths the geometric features of the scene, and the point cloud image is generally red. The mean filtering (b5) preserves the geometric features of the scene, but some point clouds are also red. The MDS obtained by the bilateral normal filtering (b4) is generally good. Voxel filtering (b6) and the PCPNet method (b7) have similar denoising effects, but some outliers appear. The method of this invention (b8) overcomes the appearance of outliers, and the overall image is blue.

[0186] Table 2 below shows the MMDS obtained after denoising the working bridge-2 scene of the dam using six methods, with standard deviations of 0.5%, 1%, and 2% of the diagonal length of the point cloud bounding box.

[0187] Table 2. MMDS results of denoising in the scene of the working bridge-2 of the river dam.

[0188] ;

[0189] As shown in Table 2, the method of this invention exhibits strong denoising capabilities in the scenario of the working bridge-2 of the river dam. When the Gaussian noise standard deviation is 0.5%, the MMDS of the method of this invention is lower than other methods. For example, the error of median filtering is as high as 43.31224 mm. Voxel filtering and PCPNet perform reasonably well under low noise conditions (0.07483 mm and 0.07484 mm, respectively), approaching the performance of the method of this invention.

[0190] As the noise level increases, at a noise level of 2%, the MMDS of voxel filtering (0.69121 mm) and the PCPNet method (0.69071 mm) are both approximately 2.5 times that at 1%, while the MMDS increase of the method of this invention is smaller. This difference indicates that voxel filtering and the PCPNet method are prone to failure in high-noise environments, while the method of this invention can effectively cope with complex noise. The errors of median filtering and mean filtering remain high throughout the noise range, indicating that they have poor adaptability to noise levels in complex scenes. In addition, under 2% Gaussian noise, the MMDS of the median filtering method for the working bridge-2 is the highest among the eight typical components discussed in this invention, and the MMDS of the method of this invention also reaches its maximum in this scenario. Combined with point cloud information analysis, this is due to the lack of point cloud information at the top of the middle crossbeam of the working bridge-2. This lack is because the scanner laser cannot reach the top of the crossbeam.

[0191] like Figure 14 The image shown is a comparison of the visualization results before and after denoising in the dam and traffic bridge-1 scene under 2% noise. Figure 14(a) is the Poisson reconstruction image, and (b) is the point cloud image. Subscripts 1-8 represent the following: clean point cloud data - 1, noisy point cloud data - 2, median filtering result - 3, bilateral normal filtering result - 4, mean filtering result - 5, voxel filtering result - 6, PCPNet result - 7, and the filtering result of the method of this invention - 8. In (b), the color of the point cloud reflects the magnitude of the MDS value; smaller values ​​are blue, and larger values ​​are red.

[0192] Depend on Figure 14 It can be seen that the overall shape of the dammed bridge-1 is straight, with wing walls and railings on both sides of the bridge deck. The Poisson reconstructed image can reflect the quality of the denoising algorithm from these parts. The median filtering (a3) ​​and normal bilateral filtering (a4) images show that the wing walls and railings are blended together, excessively smoothing the point cloud information between the two components. Mean filtering (a5) is relatively coarse in its processing of the point cloud at the corner of the wing wall. The denoising effects of voxel filtering (a6) and the intelligent algorithm PCPNet (a7) are visually similar, but the noise points at the wing walls in the lower half of the image are not removed by either method. The method of this invention (a8) removes the noise points of the wing walls while retaining the information of the railings and corners.

[0193] Of the eight images (b1-b8), b1 and b2 represent unprocessed clean point clouds and noisy point clouds, respectively, while the remaining six images show the results after denoising using different algorithms. Median filtering (b3) performs poorly on both sides of the bridge surface, with most point clouds appearing red, indicating a large MDS value. Bilateral normal filtering (b4) and voxel filtering (b6) have similar denoising effects, with the overall MDS value appearing green, but outliers appear at the edges. Mean filtering (b5) shows point clouds with large MDS values ​​at the top and bottom of the wing wall and some railings. The PCPNet method (b7) produces an overall blue-green point cloud, failing to remove noise outside the wing wall in the lower half of the image. The method of this invention (b8) produces an overall blue point cloud, overcoming the appearance of outliers and also removing noisy point clouds outside the wing wall.

[0194] Table 3 below shows the MMDS obtained after denoising the scene of the dam traffic bridge-1 using six methods, with Gaussian noise having standard deviations of 0.5%, 1%, and 2% of the diagonal length of the point cloud bounding box.

[0195] Table 3. MMDS results of denoising in the scene of the river dam and traffic bridge-1

[0196] ;

[0197] Table 3 shows that the data from the dammed river bridge-1 scenario further validates the superiority of the method of this invention. Under Gaussian noise with a standard deviation of 0.5%, the MMDS of the method of this invention is lower than that of other methods. For example, the error of median filtering is as high as 3.82721 mm, which is 56 times that of the method of this invention. The PCPNet method (0.10024 mm) performs reasonably well under low noise conditions, but is still about 47% higher than the method of this invention. This indicates that the method of this invention has a significant accuracy advantage under low-noise conditions.

[0198] As the noise level increases, the MMDS of the method of this invention increases only slightly. For example, at a noise level of 2%, the MMDS of the method of this invention increases to 0.19505 mm, while the MMDS of normal bilateral filtering (0.62620 mm) and voxel filtering (0.62686 mm) are approximately 2.5 times that at 1%. This difference indicates that traditional methods are prone to failure under high noise levels, while the method of this invention can effectively handle different noise levels. Normal bilateral filtering and voxel filtering perform relatively well under low noise conditions, but their MMDS increases rapidly under high noise levels, indicating that they are more suitable for handling low noise levels. The MMDS of median filtering and mean filtering remains high throughout the noise range, indicating that they have poor adaptability to complex scenes and are insensitive to changes in noise level. In contrast, the method of this invention maintains a low MMDS under both low and high noise conditions.

[0199] like Figure 15 The image shown is a comparison of the visualization results before and after denoising in the dam and traffic bridge scene-2 ​​under 2% noise. Figure 15 (a) is the Poisson reconstruction image, and (b) is the point cloud image. Subscripts 1-8 represent the following: clean point cloud data - 1, noisy point cloud data - 2, median filtering result - 3, bilateral normal filtering result - 4, mean filtering result - 5, voxel filtering result - 6, PCPNet result - 7, and the filtering result of the method of this invention - 8. In (b), the color of the point cloud reflects the magnitude of the MDS value; smaller values ​​are blue, and larger values ​​are red.

[0200] Depend on Figure 15As can be seen, the overall scene of the dam traffic bridge-2 is L-shaped, mainly composed of the ground of the traffic bridge section, wing walls, and railings. The advantages and disadvantages of each denoising method can be observed from the components of the Poisson reconstructed image. Median filtering (a3) ​​severely smooths the components, leaving only a portion of the wing walls in the image. Mean filtering (a5) excessively smooths the working bridge-2, resulting in severe distortion of the Poisson reconstructed image. Bilateral normal filtering (a4) and voxel filtering (a6) preserve the geometric features of the scene, but basically do not remove noise from the surface of the wing walls. The denoising effects of the intelligent algorithm PCPNet (a7) and the method of this invention (a8) are visually similar, and both can retain more original point cloud information.

[0201] The higher the MDS value in a point cloud image, the closer the point cloud color is to red. The clean point cloud (b1) and the noisy point cloud (b2) images are generally blue. After processing this scene using median filtering (b3), the overall MDS value gradually increases from the middle to both ends, and distortion occurs in the wing walls and railings. Mean filtering (b5) results in larger MDS values ​​at the upper and lower edges of the wing walls and at the railings, appearing red. Bilateral normal filtering (b4) and voxel filtering (b6) have similar denoising effects, but some outliers exist at the edges. The intelligent algorithm PCPNet (b7) reduces the occurrence of outliers. The method of this invention (b8) preserves geometric features while maintaining a smaller overall MDS value.

[0202] Table 4 below shows the MMDS obtained after denoising the dam traffic bridge-2 scene using six methods, with standard deviations of 0.5%, 1%, and 2% of Gaussian noise based on the diagonal length of the point cloud bounding box.

[0203] Table 4. MMDS results of denoising in the scene of the river dam and traffic bridge-2

[0204] ;

[0205] As shown in Table 4, the proposed method exhibits superior denoising capabilities in the dammed bridge-2 scenario. Under Gaussian noise with a standard deviation of 0.5%, the MMDS of the proposed method is lower than other methods. For example, the MMDS of median filtering is as high as 2.97398 mm, which is 64 times that of the proposed method. PCPNet performs reasonably well under low-level noise conditions (0.11357 mm), but is still approximately 146% higher than the proposed method. This demonstrates the significant accuracy advantage of the proposed method in low-level noise environments.

[0206] As the noise level increases to 1% and 2%, the error growth of the method of this invention is relatively small. For example, at a noise level of 2%, the MMDS of the method of this invention increases to 0.11351 mm, while the MMDS of normal bilateral filtering (0.65500 mm) and voxel filtering (0.65620 mm) are approximately 2.6 times that at 1%. This difference indicates that traditional methods are prone to failure in high-noise environments, while the method of this invention can effectively cope with this. Normal bilateral filtering and voxel filtering perform close to intelligent algorithms under low-noise conditions, but their MMDS increases rapidly under high-noise conditions, indicating that they have poor robustness to noise. The MMDS of median filtering and mean filtering remain high throughout the noise range, indicating that they have poor adaptability to complex scenes. The MMDS of intelligent algorithms all show linear growth, while the MMDS of the method of this invention remains within a small range, demonstrating its stability.

[0207] like Figure 16 The image shown is a comparison of the visualization results before and after denoising of the dam pier scene under 2% noise. Figure 16 (a) is the Poisson reconstruction image, and (b) is the point cloud image. Subscripts 1-8 represent the following: clean point cloud data - 1, noisy point cloud data - 2, median filtering result - 3, bilateral normal filtering result - 4, mean filtering result - 5, voxel filtering result - 6, PCPNet result - 7, and the filtering result of the method of this invention - 8. In (b), the color of the point cloud reflects the magnitude of the MDS value; smaller values ​​are blue, and larger values ​​are red.

[0208] Depend on Figure 16 It can be seen that the dam piers consist of the immersion surface and the working gate track. The Poisson reconstructed image can reflect the merits of each denoising algorithm. The median filter (a3) ​​and mean filter (a5) images show excessive smoothing, resulting in excessive corner distortion between the working gate track and the immersion surface. The bilateral normal filter (a4), voxel filter (a6), intelligent algorithm PCPNet (a7), and the method of this invention (a8) can preserve the point cloud information at the corners.

[0209] The larger the MDS value in a point cloud image, the closer the point cloud color is to red. The clean point cloud (b1) and the noisy point cloud (b2) are not denoised, so they are generally blue. The median filtering method (b3) results in a larger MDS value in the middle, making it red. The mean filtering method (b5) retains most of the geometric features, but some point clouds are also red. The normal bilateral filtering method (b4) results in a blue upper half, with green point clouds concentrated in the lower half of the gate pier. The voxel filtering method (b6) and the PCPNet method (b7) both produce green points overall, indicating that the MDS value is concentrated in the middle of the color band. The method of this invention (b8) retains the original information while being generally blue.

[0210] Table 5 below shows the MMDS obtained after denoising the dam pier scene using six methods under Gaussian noise with standard deviations of 0.5%, 1%, and 2% of the diagonal length of the point cloud bounding box.

[0211] Table 5. MMDS results of noise reduction for the dam pier scene.

[0212] ;

[0213] As shown in Table 5, the method of this invention performs excellently in the scenario of a river dam pier. Under Gaussian noise with a standard deviation of 0.5%, the MMDS of the method of this invention is lower than other methods. For example, the MMDS of median filtering is as high as 1.51654 mm, which is 33 times that of the method of this invention. The normal bilateral filtering performs well under low noise conditions (0.04734 mm), with a very small difference from the method of this invention. This indicates that the method of this invention still has a certain accuracy advantage under low-level noise conditions in this scenario.

[0214] As the noise level increases to 1% and 2%, the MMDS of the method in this invention shows a relatively small increase. For example, at a noise level of 2%, the MMDS of the method in this invention only increases to 0.11780 mm, while the MMDS of voxel filtering and PCPNet are approximately 2.8 times that at 1%. This difference indicates that traditional methods and PCPNet methods are prone to failure in high-noise environments, while the method in this invention can effectively cope with noise variations. Furthermore, the errors of median filtering and mean filtering vary relatively little across the entire noise range, indicating that they are not sensitive to fluctuations in noise level. Normal-biased filtering performs well under low-noise conditions, with its MMDS differing from that of the method in this invention by only 0.00148 mm, but its MMDS expands rapidly under high-noise environments, indicating that the denoising capability of this traditional algorithm is limited.

[0215] Based on a comprehensive analysis of the denoising results of the five components of the dam, among traditional algorithms, voxel filtering and normal bilateral filtering algorithms showed good denoising performance at the 0.5% noise level; normal bilateral filtering and PCPNet methods outperformed other denoising algorithms at the 1% and 2% noise levels; while the denoising method proposed in this invention achieved the best results, demonstrating better denoising performance under different noise levels.

[0216] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for denoising point clouds of a river dam based on a deep neural network for reinforcement learning decision-making, characterized in that, Includes the following steps: Step S1: Construct a multi-scale feature extraction network; The multi-scale feature extraction network includes a spatial transformation network and a feedforward neural network. Normalized point cloud data is input into the network. First, it passes through the spatial transformation network to obtain transformed features. Then, the transformed features are fused with the normalized features and subsequently input into the feedforward neural network to increase their dimensionality. After repeating the above process once, the resulting increased-dimensional features are summed and input into the feedforward neural network to obtain the final result. Step S2: Based on the multi-scale feature extraction network constructed in step S1, construct a point cloud denoising framework based on reinforcement learning decision-making. The point cloud denoising framework based on reinforcement learning decision-making includes an encoder, a point cloud denoising block, a decoder, and a path decision agent. It uses reinforcement learning decision-making to replace traditional loss and adaptively selects the best denoising path for each point in the point cloud. Step S3: Using the squared extreme distance between the ground real point cloud and the corresponding points in the denoised point cloud as evaluation metrics, verify the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision-making.

2. The method for denoising point clouds of a river dam based on a deep neural network for reinforcement learning decision-making, as described in claim 1, is characterized in that... The process of constructing the multi-scale feature extraction network in step S1 is as follows: Step S11: Construct a spatial transformation network suitable for point cloud data. The spatial transformation network includes at least a localization grid and a grid generator. The positioning grid receives input feature maps with a batch size of B, a channel of C, and a length of L. and output That is, the transformation applied to the feature map Parameters: =floc( ); The size varies accordingly based on the type of parameterized transformation: （1） In the above formula, , , This is the original information. , , For the transformed information, Transformation parameters; In formula (1), the affine transformations corresponding to the tensors of the three dimensions are given in sequence. The affine transformation corresponding to a 2-dimensional tensor of size [3, 4]. The size is [2, 3], and so on, with corresponding changes. floc() takes a fully connected or convolutional network and includes at least one final regression layer to generate the transformation parameters. ; pass The parameter group can perform rotation, scaling, and translation operations on a given input feature map. After transformation , The transformation parameters generated by the localization network are used to output features. When the output feature is defined as located in the feature Regular grid When the above is done, an output feature map is formed. This leads to a coordinate matrix transformation relationship: （2） （3） In the above formula, and These are the coordinates of the output target point cloud. and These are the coordinates of the origin cloud. Indicates an affine relationship; The mapping relationship of the coordinate matrix is ​​from the target feature to the input feature. The purpose of the grid generator is to sample the target feature on the original feature. Each time, pixels are sampled from different positions of the original feature and transferred to the target feature, thus covering the target feature. The position of the target feature is fixed in each sampling process, while the position of the sampled original feature is not fixed. Step S12: Construct a feedforward neural network. The feedforward neural network includes at least one input layer, one output layer, and multiple hidden layers. The neurons in the previous layer and the neurons in the next layer of the feedforward neural network are all fully connected. The complete training cycle of the feedforward neural network includes the following process: Initialization parameters: The weights w and biases b of the feedforward neural network are randomly initialized during the first iteration; Forward propagation: During training, the input data is propagated forward through multiple fully connected layers. The input data is linearly combined with the weights and biases of each layer, and then a non-linear activation function is used to improve the overall expressive power. Backpropagation: The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the model parameters. Starting from the output layer, the partial derivatives of the loss with respect to the parameters are calculated layer by layer, and this gradient information is propagated from the output layer back to the input layer. Step S13: Construct a residual neural network; The residual neural network is used to address the problem of decreased network expressive power, ensuring that the expressive power of a more layered network model is not lower than that of a less layered network model. The expression of the residual neural network is as follows: （4） In the above formula, This is the output vector; The input vector; This represents the residual mapping.

3. The method for denoising point clouds of a river dam based on a deep neural network for reinforcement learning decision-making, as described in claim 2, is characterized in that... The specific process of backpropagation in step S12 is as follows: Step S121: Calculate the loss; The error between the predicted output and the true label is calculated using a loss function. The mathematical expression for the loss function is: （5） In the above formula, n is the number of samples; This is the actual value; This is a predicted value; Considering that the noisy input data may contain outliers in the actual calculation process, the loss function given by formula (5) is greatly affected by outliers, which may lead to the inability to find a suitable gradient. Therefore, the average distance between the model's predicted value and the sample's true value is used as the loss function, and its expression is: （6） Step S122: Gradient calculation; The gradient of the loss function with respect to each weight and bias is calculated using the chain rule; Step S123: Weight update; Each weight and bias is updated using the gradient descent algorithm to reduce error. The calculation formula is as follows: （7） In the above formula, It is weight; L is the learning rate; L is the loss function. It is the partial derivative of the loss function with respect to the weights; Step S124: Parameter update; After obtaining the gradient, the Adam optimizer is used to update the model parameters. Adam uses a window decay accumulation of first-order momentum, expressed by the following formula: （8） In the above formula, The first-order momentum of the current step; This is the first-order momentum of the previous step; The historical first-order momentum retention rate; Adam uses a window decay accumulation of second-order momentum, expressed by the following formula: （9） In the above formula, Let be the second momentum of the current step; This is the second momentum of the previous step; The decay rate of the historical second-order momentum; The final update formula is: （10） In the above formula, This represents the amount of weight update; The learning rate; The first-order momentum of the current step; Let be the second momentum of the current step; A coefficient to increase the stability of the denominator.

4. The method for denoising point clouds of a river dam based on a deep neural network for reinforcement learning decision-making, as described in claim 1, is characterized in that... Step S2 involves constructing a point cloud denoising framework based on reinforcement learning decision-making, as detailed below: Set a point cloud set N, For point cloud collection The point cloud set after adding noise is used for point cloud denoising, which involves finding the set. and set The 3D geometric relationship of the midpoint cloud, noise set The expression is as follows: （11） In the above formula, It is a set of Gaussian noise; It is a collection of point clouds containing noise; The encoder will process noisy point clouds. Normalized to a unit sphere centered at the origin, for each point in the point cloud. Construct local regions and set different sphere query radii or use multiple layers with different resolutions to extract features; The decoder is positioned after the last denoising block, and the decoder processes the shift vector. Regression calculations are performed, and the point cloud is output after max pooling and a fully connected layer to remove noise. The point cloud denoising block includes multiple stacked denoising blocks. The denoising block is first connected to a shared multilayer perceptron, and then merged with the original features before entering the agent decision-making process. The denoising block contains three denoising paths: the first is to directly output the features; the second is a multilayer perceptron MLP; and the third is a multi-scale feature extraction network. An agent based on reinforcement learning decision-making selects a path for denoising. The agent based on reinforcement learning decision-making includes a multilayer perceptron (MLP), a max pooling layer, a long short-term memory (LSTM) module, a feedforward neural network (FNN), and activation functions. Traditional deep learning algorithms determine the training direction through loss, which cannot achieve ideal results in path selection. Therefore, reinforcement learning decision-making is used to replace traditional loss. Suppose the state of the given reinforcement network is Network actions Network intelligent agents are The relationship between the three is as follows: （12） In the above formula, For intelligent agents; To provide status based on the network environment The action made at that time; Based on the above, further network rewards will be introduced. The initial noise level is incorporated into the reward to guide the agent to pay more attention to noise information; network reward The formula is expressed as: （13） In the above formula, This is the action in reinforcement learning, and its value is either 0 or 1. Penalty parameters for path selection; This represents the loss between the original noise and the clean data; The loss between the denoised results and the clean data; Noise level; For scaling parameters; These are the basic weight parameters.

5. The method for denoising point clouds of a river dam based on a deep neural network for reinforcement learning decision-making, as described in claim 1, is characterized in that... In step S3, the maximum-minimum distance squared (MDS) and the mean-minimum distance squared (MMDS) between corresponding points in the ground truth point cloud and the denoised point cloud are expressed by the following formulas: （14） （15） In the above formula, It is a real point cloud; It is a point cloud after noise reduction; These are points in a real point cloud; These are the points in the point cloud after denoising. Represents the corresponding point cloud after denoising of Nearest neighbor points, Set to 10; This represents the actual number of point clouds on the ground; both MDS and MMDS are measured in mm. The smaller the values ​​of MDS and MMDS, the better the denoising performance of the constructed point cloud denoising framework based on reinforcement learning decision-making.

Images