Point Cloud Data Hole Repair Aggregation Algorithm Based on Optimized Neural Network

In order to solve the problem of cost cloud data and hole repair e ﬃ ciency and accuracy, this article o ﬀ ers a study of integrated cloud network hole algorithm research based on optimal neural network. This paper ﬁ rst introduces a common point cloud hole- ﬁ lling algorithm, provides a neural network-based point cloud blank ﬁ lling algorithm, and introduces hotspot problems in a given algorithm, combined with the technology related to high-performance computing, the parallel optimization technology based on OpenMP and CUDA is adopted to accelerate the algorithm accordingly. Experiments have shown that the accuracy of pre- and postoptimization of CUDA-based algorithms varies. As the model and input point cloud data increase, the accuracy of the algorithm decreases slightly. However, when the data increases to 104 orders of magnitude, the rate of accuracy decline meets an in ﬂ ection point and ﬁ nally stabilizes to about 96.8% in the environment of 132457 data of monk model. The point cloud hole- ﬁ lling algorithm based on the optimal neural network given in this article is highly predictable, can well repair incomplete point cloud holes, has good repair e ﬀ ect on point cloud holes, and can obtain high acceleration ratio, which can provide reference for practical engineering application.


Introduction
Chinese ancient architecture is an important witness of Chinese history and represents the inheritance and development of Chinese architecture. It is the bounden task of modern people to protect this kind of architecture, that is, to protect Chinese history. With the rapid development of laser surveying and mapping, computer virtual, image processing, 3D modeling, and other technologies, 3D laser scanning technology and virtual reality technology have also been greatly improved. They have been widely popularized and applied in the fields of ancient building reconstruction, mold manufacturing, 3D printing, and other fields and achieved good results [1]. There is no point cloud data due to the model itself being damaged during the 3D point cloud data collection, or the laser scanned line of sight being blocked, which directly affects the modeling quality. Therefore, in order to make the modeling appear smooth, it is necessary to repair the hole. The hole patching algorithm of point cloud data has been improved and developed, but there are many unclosed holes in the measured point cloud data due to the defects of measuring objects and instruments, complex measuring environment, and other factors. 3D reconstruction technology is a subject mixed intersection technology, which is widely used in pattern recognition, virtual reality, and other fields [2]. Among them, threedimensional reconstruction methods are mainly divided into interpolation-based reconstruction methods and triangular mesh-based reconstruction methods. The research of such methods is becoming more and more mature, but there are still some shortcomings, which makes the effect and speed of reconstruction become a bottleneck. To improve the speed of recovery, the dot cloud design gap needs to be repaired quickly; the existing mature parallel technology can be used to optimize the point cloud hole repair algorithm in parallel and quickly improve the efficiency of hole repair [3,4].
In the field of machine learning and cognitive science, artificial neural networks are similar in structure and function to the human brain. It is a central neural network system with adaptive nonlinear characteristics composed of a large number of neurons, a mathematical or computational model commonly used to estimate or approximate functions. Artificial neural network can appropriately change its composition structure according to the attributes of input information. It is an adaptive system [5]. The hidden layer is used to complete the corresponding calculation operation; that is, the operation calculation results are transferred to the output layer. The output layer is mainly used to process and display the final calculation results. Because of its unique advantages, neural network has good applications in many fields, especially in classification and recognition. The neural network interprets the data input by the sensor through the machine sensing system, marks the original data, and clusters it. The pattern recognized by neural network is mainly based on the vector of array type. Therefore, before processing the data, we need to convert the data into numerical value and then operate the converted data. Neural network can realize clustering and classification operations, and it can be thought of as a cluster or classification layer at the top of the stored and managed data. For unlabeled data, neural networks can group data according to the similarity of the input sample. Therefore, the deep neural network can be regarded as a very important part of the machine learning system, which also includes reinforcement learning, classification, and regression algorithms [6]. Figure 1 shows the process of optimizing the neural network to restore the 3D mesh surface.

Literature Review
In recent years, in-depth study has been well used in image recognition and targeting, which has attracted extensive attention, and has shown good application prospects in other fields. Therefore, it has become a research point of extensive concern in various fields. Neural network in deep learning is the focus of research. When neural network is used to operate the image, its effect is better than traditional methods, the accuracy of image recognition is higher, and the operation process is relatively simple [7]. Trybaa et al. calculate the data points through the weighted average method and take the inverse data of the node distance as the weight function and interpolate it, so as to make the object surface clear. Although this method is quite simple, effective, and widely used, when the amount of point data is large, the amount of calculation will increase and the number of calculation errors will also increase [8]. Gao et al. described a new analysis method for representing irregular surfaces, which involves the summation of quadric surface equations with unknown coefficients. The quadric surface is located at the important point of the whole area to be drawn. It is given to solve the polygonal surface equation based on the polyhedron equation and compare the polygonal surface with the irregular surface [9]. Tu et al. fitted the three-dimensional surface by improving the radial basis function and evaluated the time, storage, and accuracy by using the scattering data interpolation and testing methods, so as to improve the visual observability of the object surface, and the repair speed and accuracy have good effects [10]. Chen et al. studied the application of neutral network algorithm in surface reconstruction of scattered point cloud and implicit surface meshing. It is found that the most suitable neutral network type is the modified version of the growth cell structure. The algorithm randomly extracts a target space, usually a point cloud or an implicit surface, and adjusts the connectivity of the neutral mesh accordingly. The algorithm can better deal with sharp and concave surfaces, and the speed of the algorithm is independent of the size of the input data. It is more suitable for reconstructing surfaces from larger point sets [11]. Lin et al. presented a surface reconstruction algorithm with multiple attributes, which mainly studies multiscale scattered point cloud data. The algorithm uses the balanced binary tree to divide the object space to form several small spatial regions and then uses the radial basis function to calculate the difference, so as to integrate the repaired small regions into a complete three-dimensional model [12]. Martinez et al. found that triangulation is more suitable for cavity repair of 3D model than other methods. The point is that there are certain limitations according to the discrete relationship of cloud and noise information, there is a topologically consistent relationship between the data, and the points are given by the cloud removal algorithm and the noise [13]. Biundini et al. proposed a new method to solve three-dimensional shape deformation, which is the process of maximizing the similarity between two shapes by representing the interpolation of two shapes as the interpolation of one shape. The process incrementally optimizes the objective function and deforms the implicit surface model. The deformable surface is represented as a level set of densely sampled scalars. Since this method adopts the prediction method, there are certain uncertainties and limitations [14]. In order to reconstruct the model, Hsu and Zhuang used coordinate measuring machine or laser scanner to collect point cloud data and then converted the collected data into a model. The reconstruction of model surface was realized by neural network approximation method. Three four-layer neural networks were designed, and the designed neural network was trained by error back propagation method [15]. For the grid method in surface reconstruction, reconstruction, simplification, and stratification of 3 D grids, Duanmu and Xing proposed the algorithm for hierarchical reconstruction of surface grids with large data quantity [16].
Based on this research, this paper proposes a point cloud data hole repair aggregation algorithm based on optimized neural network. The optimized neural network is not only suitable for processing large-scale and noisy data but also the scientific community has proved that the optimization function with a hidden layer has the best approximation ability. Wireless Communications and Mobile Computing connection methods between neurons in the network system structure, it is often divided into the following types:

Research Methods
(1) Feedforward Neural Network. Feedforward neural network is also called forward neural network. The feedback signal will be generated in the training process. Because there is no feedback between the network layers, the data can be transmitted to the output layer during classification; it can only be transmitted forward in one direction, which is referred to as feedforward network for short. The network can be represented by N, the input vector is represented by X, the connection weight vector of each layer in the network is W 1~W3 , and the activation function can be represented by F 1~F3 . The network output value is O 1 = F 1 ðXW 1 Þ, the output corresponding to the network hidden layer is (2) Feedback Neural Network. A feedback neural network is a neural network with its own internal or global response from output to input. In this type of network, each neuron feeds back its output signal to other neurons as input signal. After working for a period of time, the network can reach a relatively stable state. Compared with feedforward neural network, it has higher complexity [17,18].

Topology and Network
Structure. The neural network topology is a three-layer network with one hidden layer. The function responds locally to the input signal. If the input signal x of the function is close to the center of the function, the nodes of the hidden layer produce a larger output. You can see that this network has local proximity capabilities [19]. The value r of hidden layer neurons can be obtained by the following formula: (1) Gaussian function (Gaussian function) (2) Reflected sigmoidal function (abnormal S-type function) (3) Inverse multiquadric (inverse distortion correction function) Through the analysis of several functions, it is found that Gaussian function is more suitable for the hidden layer transformation function in this study, and its expression is as follows: In the formula, φ j is the output of the j-th unit of the hidden layer; X = ðx 1 , x 2 ,⋯,x T Þ is the input vector; k:k represents a norm, usually taken as c j is the center of the j Gaussian unit in the hidden layer; E is n × 1 unit vector; δ j is the radius.

Wireless Communications and Mobile Computing
The output of a neural network can be expressed as follows.

Implementation of Point Cloud Data and Hole Repair
Algorithm of Neural Network. Main steps of point cloud filtering and hole repair using neural network are the following: (1) Collection and normalization of prediction data and training data: silent point cloud data is extracted from raw cloud data loaded in MATLAB R2016a, part of the data is randomly selected as the prediction data, the rest is the training data, and then, the two parts of the data are normalized between [0,1]  Figure 2.
The point cloud hole repair algorithm uses the scattered point cloud data points in the point cloud model as the basis of the research, and the amount of data generated is usually used, hundreds of millions, and each data has a certain independence for a very large number of data points can be combined with the corresponding parallel technology. For a very large number of data points, the corresponding parallel technology can be combined to perform parallel operations on a large number of data points. Read the point cloud data from the entry point cloud model and use the octal data. This structure stores the point cloud data collected in the point cloud model, and the stored information mainly includes the coordinates of each data point in the point cloud model, number of octet layers and suboctet pointers, and operation of octree data structural elements.

CUDA-Based Point Cloud
Hole-Filling Algorithm. The programming model based on GPU (Compute Unified Device Architecture, CUDA) has strong floating-point operation ability. When the amount of calculation is large, the calculation advantage of GPU is more obvious, and it has the advantages of better multithreaded calculation ability and less transmission delay.
Pay attention to the following points when using CUDA programming: initialize the video memory at the GPU end. When transmitting a large amount of data, the transmission delay between GPU and CPU becomes higher and higher, which reduces the transmission efficiency of data. In order to transmit data quickly and effectively, the transmission frequency between GPU and CPU can be reduced [20].
This chapter adopts CUDA parallel technology based on GPU to optimize the point cloud hole repair algorithm accordingly. When using convolutional neural network to repair the point cloud hole, there are three main parts with large computational overhead, and parallel optimization is carried out for the parts with hot spots (for the process of  Wireless Communications and Mobile Computing using CUDA to calculate surface parameter information in parallel, see Figure 3).

Analysis of Parallel Optimization Results Based on
OpenMP. The previous section analyzes and compares the serial average running time of algorithms on different platforms through experiments. In this section, OpenMP based on CPU is used to optimize the hot issues in the algorithm.  Figure 4. According to the above figure, when OpenMP is used for optimization, for the same point cloud model, the more threads the algorithm allocates, the less the average running time of the algorithm. For different point cloud models, the average run time of the algorithm also increases as the amount of point cloud data contained in the point cloud model increases. For the ball model, due to the small amount of point cloud data contained in the model, when using multithreading to execute the model, the running cost of threads is large. Therefore, as the number of circuits increases, so does the algorithm's run time. When the point cloud model contains more point cloud data or the number of threads allocated in the algorithm is more, the shorter the algorithm, the lower the algorithm's performance.
In addition, this paper makes a corresponding comparative analysis on the algorithm running time of different point cloud models to allocate different number of threads under platform 2. The specific experimental results are shown in Figure 5.
It can be seen from the above figure that when OpenMP is used for optimization, when a single point cloud model contains a large amount of point cloud data in a point cloud     Table 1 and Figure 6.
As can be seen from the figure and table above, the different models of point clouds on platform 3 increase the number of circuits and the algorithm run time. For the same number of circuits, the higher the point cloud data contained in the point cloud model, the shorter the average algorithm run time. The trend is more pronounced. It can be seen that the average operating time of the algorithm is clearly related to the amount of data contained in the point cloud model and the number of circuits.

Comparison of Algorithm Accuracy under Different
Platforms. This section mainly analyzes and compares the accuracy of the algorithm based on OpenMP before and after optimization. This paper compares the number of preoptimization point cloud data and the postoptimization point cloud data on different platforms to determine the accuracy of the algorithm before and after optimization. The specific situation of the number of point cloud data before and after optimization is shown in Table 2, and the accuracy before and after algorithm optimization based on OpenMP is shown in Table 3.
According to Tables 2 and 3, the change of the number of point clouds is within the normal error range and has a certain controllability. Before and after the optimization of the algorithm, the number of point cloud data in the point cloud model does not change much but increases or decreases slightly accordingly. When the platform is more complex, the change range of the number of point clouds increases slightly, but it is generally within the error range. Therefore, the accuracy of the algorithm does not change much before and after optimization.

Analysis of Parallel Optimization Results
Based on CUDA. In this paper, NVIDIA Tesla c2070 graphics card is used for experimental analysis, and the parallel efficiency and speedup ratio of the algorithm is compared through experiments. Different point cloud models are tested under platform 3. The specific test results are shown in Table 4.     Figure 7. According to Figure 7, the running time of different point cloud models running on platform 3 after parallel optimization is significantly lower than that of serial algorithm. Therefore, parallel optimization based on GPU has certain feasibility and effectiveness. According to Figure 8, under the same experimental platform, when using CUDA for parallel optimization, the more point cloud data in the point cloud model, the higher the acceleration ratio of the algorithm.

Comparison of Algorithm Accuracy under Different
Platforms. This section mainly analyzes and compares the accuracy before and after CUDA-based algorithm optimization. Among them are the main points of analysis of serial algorithms and number of cloud data points and cloud models downloaded on platform 1, platform 2, and platform 3 using CUDA parallel optimization number of point cloud data, as shown in Table 5.
Before and after optimizing the algorithm, the number of point cloud data in the point cloud model does not change much, as Table 6 shows that the point cloud data for that model may be missing to optimize the object model. It can be seen from Table 5 that the number of point cloud data in the point cloud model before and after optimization of the algorithm does not change much, because when optimizing the object model, the point cloud data of the model may be missing due to human reasons, but the point cloud data has only slight increase or less, and there is no widespread or massive data loss. As can be seen from Table 5, there are clear changes in the accuracy of the CUDA-based algorithm before and after optimization. Since the model contains more dot cloud data, the algorithm's accuracy      7 Wireless Communications and Mobile Computing decreases slightly, but the data increases in the order of 104 points, and the accuracy rate decreases to about 96.8% in a monkey model environment with a data size of 132,457 when a bending point occurs. In addition, the accuracy of algorithms that work on different platforms varies for the same point cloud model. As the complexity of the platform increases, the accuracy of the algorithm decreases slightly, but in the point cloud model, the precision of point cloud data or different action programs and algorithms the accuracy has hardly changed. Therefore, the accuracy of the algorithm before and after the optimization is almost the same, and there is not much difference.

Conclusion
For the scattered point cloud data collected in this document, the octree data structure was decomposed into the entire point cloud data, and then, the hierarchical operation of the point cloud model was completed. The twodimensional coordinates in three directions are obtained through spatial mapping. Use neural networks to unravel the hole features of point cloud models. According to the extracted hole features, the corresponding sparse representation classification model is constructed, and the residuals of each category are calculated to generate the boundary of the point cloud hole, obtain the pure hole area, and then complete the filling of the polygon hole. When using a multilayer CNN convulsive neural network to fill a point cloud gap, the efficiency of the algorithm is low due to factors such as the size of the network and the complex operational process. The point cloud hole compensation algorithm, which was aimed at solving a large number of scattered cloud data computational problems, optimizes a combination of commonly used parallel computing technologies, such as CPU acceleration and GPU acceleration technology, to quickly reduce downtime, improve algorithm performance, and increase algorithm efficiency.

Data Availability
No data were used to support this study.

Conflicts of Interest
The author declares that there are no conflicts of interest regarding the publication of this article.