Deep learning based multi-beam forward-looking three-dimensional sonar data compression method and system

Abstract

The application discloses a kind of multi-beam forward-looking three-dimensional sonar data compression method and system based on deep learning, comprising: in three-dimensional sonar data compression lower computer platform, after three-dimensional sonar data is carried out reflection intensity normalization and point cloud block, it is obtained by first deep learning network including first MLP layer, first Attention layer and first Point Transformer layer and is reduced sampling, compressed three-dimensional sonar data transmission is obtained to three-dimensional sonar data decompression upper computer platform;In three-dimensional sonar data decompression upper computer platform, after the compressed three-dimensional sonar data received is carried out up sampling by second deep learning network including second MLP layer, second Attention layer and second Point Transformer layer, intensity data recovery is carried out, and three-dimensional sonar data is obtained after decompression and restoration.The application can realize high-quality and efficient three-dimensional sonar data compression, transmission and decompression.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of three-dimensional sonar, and particularly relates to a multi-beam forward-looking three-dimensional sonar data compression method and system based on deep learning. BACKGROUND

[0002] In the vast field of ocean exploration, the rise of multi-beam forward-looking three-dimensional sonar technology marks a major leap in underwater exploration technology. This technology not only greatly broadens the cognitive boundaries of mankind in the unknown world of the ocean, but also plays an irreplaceable role in multiple key areas such as underwater topographic mapping, marine environment monitoring, and seabed resource exploration. With the progress of technology and the growing demand, multi-beam forward-looking three-dimensional sonar systems are gradually developing towards higher precision, faster response, and stronger real-time performance, which cannot be achieved without effective processing and efficient transmission of massive sonar data.

[0003] Multi-beam forward-looking three-dimensional sonar systems are usually divided into two main modules, namely the lower computer for collecting and processing sonar data and the upper computer for displaying sonar data, which interact through network cables or wireless networks. Compared with traditional scanning sonars, multi-beam forward-looking three-dimensional sonars can provide more real-time data, achieving a scanning rate of one frame per second or higher, without the need for waiting for the delay of stitching beam images. At the same time, compared with multi-beam two-dimensional forward-looking sonars, multi-beam three-dimensional forward-looking sonars need to detect data in an additional dimension, providing high-resolution three-dimensional modeling data that is more clear and intuitive than two-dimensional plane data. However, the resulting large amount of data poses a serious challenge to computing and transmission capabilities. In order to achieve real-time data processing and transmission, effective data compression algorithms and systems are urgently needed.

[0004] Currently, multi-beam forward-looking three-dimensional sonar data compression algorithms and systems are important content in the field of three-dimensional sonar research, and there are many specific compression methods, but there are the following problems:

[0005] (1) Compression algorithm efficiency and effect coexistence problem: the widely used compression algorithm based on binary stream serialization, although simple and easy to implement, often fails to achieve the desired compression ratio, especially for three-dimensional sonar data with specific distribution characteristics, the compression effect is more limited, and the compression rate is low, while some complex compression techniques can improve the compression rate to a certain extent, but the utilization rate of hardware is low, which may affect real-time performance;

[0006] (2) Data sparsity and recovery quality contradiction: three-dimensional sonar data usually has a high degree of sparsity, i.e. most areas may have no valid signals or very low signal strength, which can easily lead to distortion of the recovered data if not properly handled during compression, affecting subsequent analysis and application;

[0007] (3) The importance of reflection intensity preservation: In three-dimensional sonar data, reflection intensity is one of the important parameters reflecting the properties of underwater objects. How to effectively preserve high reflection intensity data during compression and avoid its excessive smoothing or loss during compression is the key to ensuring data quality.

[0008] Therefore, future research should focus on developing compression algorithms and systems specifically for multi-beam forward-looking three-dimensional sonar data, and intelligently optimizing the compression process to further improve compression rate and recovery quality. SUMMARY

[0009] In view of the above, the purpose of the present application is to provide a multi-beam forward-looking three-dimensional sonar data compression method and system based on deep learning, which can realize efficient and lossless point cloud compression, transmission and decompression by combining deep learning with the downsampling compression process and upsampling decompression process in three-dimensional sonar data, thereby providing strong support for application fields such as ocean exploration and underwater terrain detection.

[0010] To achieve the above-mentioned purpose of the application, the technical solutions provided by the present application are as follows:

[0011] In a first aspect, the present application provides a multi-beam forward-looking three-dimensional sonar data compression system based on deep learning, which includes a three-dimensional sonar data compression lower computer platform and a three-dimensional sonar data decompression upper computer platform.

[0012] The three-dimensional sonar data compression lower computer platform is used to normalize the reflection intensity of the acquired three-dimensional sonar data and divide the point cloud into blocks, and then perform downsampling through a first deep learning network including a first MLP layer, a first Attention layer and a first Point Transformer layer, to obtain compressed three-dimensional sonar data and transmit it to the three-dimensional sonar data decompression upper computer platform.

[0013] The three-dimensional sonar data decompression upper computer platform is used to perform upsampling on the received compressed three-dimensional sonar data through a second deep learning network including a second MLP layer, a second Attention layer and a second Point Transformer layer, and then perform intensity data recovery to obtain decompressed three-dimensional sonar data.

[0014] Preferably, the acquired three-dimensional sonar data is in the form of three-dimensional point cloud. Before inputting the first deep learning network for downsampling, the three-dimensional point cloud is first normalized for reflection intensity, and then the normalized three-dimensional point cloud is divided into blocks to obtain point cloud blocks.

[0015] Preferably, when down-sampling in the first deep learning network, the point cloud block is first subjected to farthest point sampling to obtain reserved points and deleted points, and then the features of the deleted points are extracted using the first MLP layer, the first Attention layer and the first Point Transformer layer respectively and fused to obtain aggregated information of the deleted points, and the aggregated information of the deleted points and the reserved points are constructed into compressed three-dimensional sonar data.

[0016] Preferably, the compressed three-dimensional sonar data is constructed into a data frame for transmission, and the format of the data frame includes a frame number, a maximum reflection intensity of the current frame, a minimum reflection intensity of the current frame and the compressed three-dimensional sonar data.

[0017] Preferably, the transmission of data frames between the three-dimensional sonar data compression lower computer platform and the three-dimensional sonar data decompression upper computer platform is carried out through the UDP protocol, the data frame is cut into network data packets of the UDP protocol before transmission, and then the network data packets are sent to the three-dimensional sonar data decompression upper computer platform, which identifies and merges the network data packets according to the packet header information to restore a whole frame of data frame.

[0018] Preferably, the format of the data packet includes a frame number, a packet number, a total number of packets and cut data of the compressed three-dimensional sonar data.

[0019] Preferably, when up-sampling in the second deep learning network, the aggregated information of the deleted points in the compressed three-dimensional sonar data is restored using the second MLP layer, the second Attention layer and the second Point Transformer layer respectively, the restored data is fused through the third MLP layer to obtain decompressed deleted points, and the reflection intensity data of the reserved points and the decompressed deleted points is recovered based on the maximum reflection intensity of the current frame and the minimum reflection intensity of the current frame, and finally the decompressed and restored three-dimensional sonar data is obtained.

[0020] In a second aspect, to achieve the above-mentioned object, the embodiment of the present application further provides a multi-beam forward-looking three-dimensional sonar data compression method based on deep learning, which is realized by using the multi-beam forward-looking three-dimensional sonar data compression system based on deep learning as described above, and includes the following steps:

[0021] After the three-dimensional sonar data obtained by the three-dimensional sonar data compression lower computer platform is subjected to reflection intensity normalization and point cloud blocking, down-sampling is carried out through the first deep learning network including the first MLP layer, the first Attention layer and the first Point Transformer layer to obtain compressed three-dimensional sonar data and transmit the compressed three-dimensional sonar data to the three-dimensional sonar data decompression upper computer platform.

[0022] The decompressed three-dimensional sonar data received by the three-dimensional sonar data decompression host computer platform is up-sampled by a second deep learning network including a second MLP layer, a second Attention layer and a second Point Transformer layer, and then intensity data recovery is performed to obtain decompressed three-dimensional sonar data.

[0023] In a third aspect, to achieve the above-mentioned object, the embodiment of the present application further provides a deep learning-based multi-beam forward-looking three-dimensional sonar data compression device, comprising a memory and one or more processors, the memory is used to store a computer program, and the processor is used to realize the deep learning-based multi-beam forward-looking three-dimensional sonar data compression method when the computer program is executed.

[0024] In a fourth aspect, to achieve the above-mentioned object, the embodiment of the present application further provides a computer-readable storage medium, and the storage medium stores a computer program, and the computer program is executed by a computer to realize the deep learning-based multi-beam forward-looking three-dimensional sonar data compression method.

[0025] Compared with the prior art, the present application has at least the following beneficial effects:

[0026] The present application constructs a down-sampling data compression and an up-sampling data decompression part of three-dimensional sonar data based on a deep learning network including an MLP layer, an Attention layer and a Point Transformer layer, and applies the deep learning network to a three-dimensional sonar data compression lower computer platform and a three-dimensional sonar data decompression host computer platform, realizes data acquisition, compression, transmission and decompression functions of the upper and lower computers, can simultaneously guarantee quality and efficiency of point cloud compression, and effectively retains reflection intensity information after decompression. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0028] Figure 1 is a structural schematic diagram of the deep learning-based multi-beam forward-looking three-dimensional sonar data compression system provided by the embodiment of the present application;

[0029] Figure 2 is a flowchart of down-sampling provided by the embodiment of the present application;

[0030] Figure 3is a three-dimensional sonar data frame format diagram provided by an embodiment of the present application;

[0031] Figure 4 is a three-dimensional sonar network data packet format diagram provided by an embodiment of the present application;

[0032] Figure 5 is a process schematic diagram for upsampling based on a second deep learning network provided by an embodiment of the present application;

[0033] Figure 6 is a process schematic diagram of a deep learning-based multi-beam forward-looking three-dimensional sonar data compression method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0035] The inventive concept of the present application is that, in order to solve the problems of low compression efficiency, poor data quality and loss of reflection intensity in the prior art multi-beam forward-looking three-dimensional sonar data compression algorithm and system, an embodiment of the present application provides a deep learning-based multi-beam forward-looking three-dimensional sonar data compression method and system, which respectively constructs a three-dimensional sonar data downsampling data compression and upsampling data decompression part based on a deep learning network including an MLP (Multilayer Perceptron, multilayer perceptron) layer, an Attention (attention) layer and a Point Transformer (point cloud transformer) layer, and is applied to a three-dimensional sonar data compression lower computer platform and a three-dimensional sonar data decompression upper computer platform. Through five steps including reflection intensity normalization processing, point cloud blocking, downsampling, upsampling and intensity data recovery, the data acquisition, compression, transmission and decompression functions of the lower and upper computers are realized, the quality and efficiency of the point cloud compression are ensured, and the reflection intensity information is effectively preserved after decompression.

[0036] Figure 1 is a structure schematic diagram of a deep learning-based multi-beam forward-looking three-dimensional sonar data compression system provided by an embodiment of the present application. As Figure 1As shown, the embodiment provides a deep learning-based multi-beam forward-looking three-dimensional sonar data compression system, which comprises a three-dimensional sonar data compression lower computer platform and a three-dimensional sonar data decompression upper computer platform, and the upper and lower computers communicate through a network. Through the three-dimensional sonar data compression lower computer platform, deep compression of three-dimensional sonar data can be realized at low cost. After data compression of the obtained uncompressed data, the compressed data is divided into network data packets of User Datagram Protocol (UDP) and sent to the upper computer platform. Through the three-dimensional sonar data decompression upper computer platform, the network data packets can be obtained, spliced and decompressed, and the decompressed data is sent to the display and storage program for display and saving.

[0037] Specifically, the three-dimensional sonar data compression lower computer platform is used to normalize the reflection intensity of the obtained three-dimensional sonar data and divide the point cloud into blocks, and then the first deep learning network comprising a first MLP layer, a first Attention layer and a first Point Transformer layer is used for downsampling to obtain compressed three-dimensional sonar data and transmit the data to the three-dimensional sonar data decompression upper computer platform.

[0038] In the embodiment, the obtained three-dimensional sonar data exists in the form of three-dimensional point cloud, and each point is composed of horizontal axis coordinate (x), vertical axis coordinate (y), vertical axis coordinate (z) and reflection intensity (intensity). Since the point cloud needs to be input into the first deep learning network for downsampling compression in the subsequent process, and the data of the reflection intensity is large, the three-dimensional point cloud needs to be normalized in reflection intensity. The calculation formula of normalization is as follows:

[0039]

[0040] Wherein, P is the reflection intensity value of the current point, P min and P max are the minimum value of the reflection intensity of the current frame and the maximum value of the reflection intensity of the current frame respectively, and P' is the normalized reflection intensity value of the current point. Through the above normalization method, the reflection intensity can be within the range of 0-1.

[0041] In the embodiment, after the reflection intensity normalization is completed, the normalized three-dimensional point cloud is subjected to a block operation. The block operation is to sample small point cloud blocks using KNN (K-Nearest Neighbors) algorithm, and then the point cloud blocks are input into the first deep learning network for downsampling.

[0042] In the embodiment, as shown in Figure 2As shown, down-sampling is performed based on the first deep learning network. First, for the above-mentioned completed block point cloud, the key points in the point cloud block are obtained according to the reflection intensity, and in the embodiment, the first 50% of the points in the point cloud block are obtained as the key points according to the reflection intensity. Then, a random seed is selected to perform farthest point sampling on the key points to obtain reserved points and deleted points, so that the skeleton of the point cloud data can be better collected. After completing the farthest point sampling, each reserved point and the deleted point in the current block is sent to the first deep learning network for operation, which includes a first MLP layer, a first Attention layer and a first Point Transformer layer.

[0043] In the first MLP layer, the dimension of the input data is K*4, where K is the number of deleted points in the point cloud block, and 4 corresponds to the dimension of each data point. The dimension of the output data after the first MLP layer is D*1, where D is the output channel number and 1 is the dimension of the output data. In the first Attention layer, the dimension of the input data is N*4, where N corresponds to the number of all points (including reserved points and deleted points) in the current point cloud block. The scaling dot product operation is performed in an Attention block only using the query parameters of each reserved point and the key parameters of each deleted point. The dimension of the data after the scaling dot product operation is K*4, and then the dimension of the output data is D*1 after dimension reduction by an MLP layer. In the first Point Transformer layer, the dimension of the input data is K*4, where K is the number of deleted points in the point cloud block. After a Point Transformer block and an MLP layer for dimension reduction, D*1 data is obtained. The data obtained by the first MLP layer, the first Attention layer and the first Point Transformer layer is combined into a D*1 matrix using matrix addition operation, and the aggregation information of the current deleted point is obtained while the information of the reserved point is retained. The aggregation information of the deleted point and the reserved point are constructed into compressed three-dimensional sonar data.

[0044] As shown in Figure 3 The three-dimensional sonar data frame format diagram is shown. The compressed three-dimensional sonar data is constructed into a data frame for transmission. The format of the data frame includes frame number, maximum reflection intensity of the current frame, minimum reflection intensity of the current frame and compressed three-dimensional sonar data. The frame number is used to uniquely identify a frame of data information. The maximum reflection intensity of the current frame and the minimum reflection intensity of the current frame are used to restore the reflection intensity information in the current frame. The compressed three-dimensional sonar data is sequentially composed of the reserved point data and the deleted point aggregation information in each block.

[0045] As shown in Figure 4The diagram illustrates the data packet format of a 3D sonar network. The 3D sonar data compression lower-level platform and the 3D sonar data decompression upper-level platform transmit data frames via UDP (User Datagram Protocol). Before transmission, the data frame is segmented into fixed-length UDP network data packets of no more than 65536 bytes, with packets shorter than 65536 bytes padded with zeros at the end. Each data packet includes a frame number, packet number, total number of packets, and the segmented data of the compressed 3D sonar data. The network data packets are then sent to the 3D sonar data decompression upper-level platform. The platform retrieves these network packets, reassembles them according to their packet numbers, and reconstructs a compressed 3D sonar data frame.

[0046] Specifically, the 3D sonar data decompression host computer platform is used to upsample the received compressed 3D sonar data through a second deep learning network including a second MLP layer, a second Attention layer, and a second Point Transformer layer, and then restore the intensity data to obtain the decompressed and restored 3D sonar data.

[0047] In the embodiments, such as Figure 5 As shown, upsampling is performed based on a second deep learning network. The upsampling is designed according to the principle of symmetry with downsampling. For the currently received compressed 3D sonar data frame, the aggregated information of the retained points and deleted points is fed into three networks for processing, namely the second MLP layer, the second Attention layer, and the second Point Transformer layer.

[0048] In the second MLP layer, the dimension of the input data is D*1, that is, the deletion point aggregation information of each block, and the dimension of the parameter output by the second MLP layer is K*4. In the second Attention layer, the dimension of the input data is D*1, which is first upgraded to a data matrix of K*4 through an MLP, and then a scaling dot product operation is performed on the query parameter of each reserved point and the key parameter of each deletion point through an Attention block. The dimension of the data after the scaling dot product operation is K*4. In the second Point Transformer layer, the dimension of the input data is D*1, which is first upgraded to a data matrix of K*4 through an MLP, and then the data is processed through a Point Transformer block to obtain K*4 data. The data obtained through the three networks of the second MLP layer, the second Attention layer and the second Point Transformer layer is uniformly input into a third MLP layer with an input dimension of K*12 and an output dimension of K*4 for fusion to obtain decompressed deletion points. The reflection intensity data of the reserved points and the decompressed deletion points is recovered based on the maximum reflection intensity of the current frame and the minimum reflection intensity of the current frame, and finally the three-dimensional sonar data after decompression and restoration is obtained.

[0049] In summary, the embodiment of the present application provides a multi-beam forward-looking three-dimensional sonar data compression system based on deep learning. The deep learning network structure is constructed according to the special properties of three-dimensional sonar data. The deep learning network including MLP layer, Attention layer and Point Transformer layer is used to construct the down-sampling data compression and up-sampling data decompression parts of three-dimensional sonar data, respectively. The three-dimensional sonar data compression and data decompression algorithm with low cost is realized. The data acquisition, compression, transmission and decompression functions of the upper and lower computer platforms are realized, which can guarantee the quality and efficiency of point cloud compression at the same time, and effectively retain the reflection intensity information after decompression. This is crucial for subsequent point cloud analysis, environment modeling and target recognition tasks. It provides strong technical support for the application of three-dimensional sonar data in complex environments, and can be widely used in underwater topographic mapping, marine environment monitoring and seabed resource exploration fields, greatly improving the real-time performance and accuracy of data processing.

[0050] Based on the same inventive concept, as Figure 6 As shown in the above-mentioned based on deep learning multi-beam forward-looking three-dimensional sonar data compression system, the embodiment of the present application also provides a multi-beam forward-looking three-dimensional sonar data compression method based on deep learning, which is realized by using the above-mentioned based on deep learning multi-beam forward-looking three-dimensional sonar data compression system, including the following steps:

[0051] S1, using the three-dimensional sonar data compression lower computer platform, the acquired three-dimensional sonar data is normalized and point cloud is blocked, and then is down-sampled through the first deep learning network including the first MLP layer, the first Attention layer and the first Point Transformer layer, to obtain the compressed three-dimensional sonar data and transmit to the three-dimensional sonar data decompression upper computer platform.

[0052] S2, using the three-dimensional sonar data decompression upper computer platform, the received compressed three-dimensional sonar data is up-sampled through the second deep learning network including the second MLP layer, the second Attention layer and the second Point Transformer layer, and then the intensity data is recovered to obtain the decompressed three-dimensional sonar data.

[0053] Based on the same inventive concept, the embodiment of the present application also provides a multi-beam forward-looking three-dimensional sonar data compression device based on deep learning, comprising a memory and one or more processors, the memory is used to store a computer program, and the processor is used to realize the multi-beam forward-looking three-dimensional sonar data compression method based on deep learning when the computer program is executed.

[0054] Based on the same inventive concept, the embodiment of the present application also provides a computer readable storage medium, the storage medium stores a computer program, when the computer program is executed by a computer, the multi-beam forward-looking three-dimensional sonar data compression method based on deep learning is realized.

[0055] It should be noted that the multi-beam forward-looking three-dimensional sonar data compression method based on deep learning, the multi-beam forward-looking three-dimensional sonar data compression device based on deep learning and the computer readable storage medium provided by the above embodiment all belong to the same inventive concept as the multi-beam forward-looking three-dimensional sonar data compression system based on deep learning, and the specific implementation process is detailed in the multi-beam forward-looking three-dimensional sonar data compression system based on deep learning, which will not be repeated here.

[0056] The above specific embodiments have been described in detail to explain the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application, and is not used to limit the present application. Any modification, supplement and equivalent replacement within the principle range of the present application should be included in the protection range of the present application.

Claims

1. A deep learning based multi-beam forward-looking three-dimensional sonar data compression system, characterized in that, The application relates to a three-dimensional sonar data compression lower computer platform and a three-dimensional sonar data decompression upper computer platform. The three-dimensional sonar data compression lower computer platform is used for carrying out reflection intensity normalization and point cloud blocking on acquired three-dimensional sonar data, carrying out down-sampling on the three-dimensional sonar data by a first deep learning network comprising a first MLP layer, a first Attention layer and a first Point Transformer layer, obtaining compressed three-dimensional sonar data and transmitting the three-dimensional sonar data to the three-dimensional sonar data decompression upper computer platform. The three-dimensional sonar data decompression upper computer platform is used for carrying out intensity data recovery on the received compressed three-dimensional sonar data by a second deep learning network comprising a second MLP layer, a second Attention layer and a second Point Transformer layer after up-sampling, and obtaining decompressed three-dimensional sonar data. The acquired three-dimensional sonar data is in the form of three-dimensional point cloud, and before being input into the first deep learning network for down-sampling, the three-dimensional point cloud is first subjected to reflection intensity normalization, and then the normalized three-dimensional point cloud is subjected to blocking to obtain point cloud blocks.

2. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 1, wherein, In the first deep learning network, the point cloud blocks are first subjected to farthest point sampling to obtain reserved points and deleted points, and then the first MLP layer, the first Attention layer and the first Point Transformer layer are used to respectively extract features of the deleted points and fuse the features to obtain aggregated information of the deleted points, and the aggregated information of the deleted points and the reserved points are constructed into compressed three-dimensional sonar data.

3. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 2, wherein, The compressed three-dimensional sonar data is constructed into a data frame for transmission, and the format of the data frame comprises a frame number, a maximum reflection intensity of a current frame, a minimum reflection intensity of the current frame and the compressed three-dimensional sonar data.

4. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 3, wherein, The three-dimensional sonar data compression lower computer platform and the three-dimensional sonar data decompression upper computer platform transmit the data frame through a UDP protocol, and before transmission, the data frame is cut into network data packets of the UDP protocol, and then the network data packets are sent to the three-dimensional sonar data decompression upper computer platform, which identifies and combines the network data packets according to the packet header information to restore the data frame.

5. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 4, wherein, The format of the data packet comprises a frame number, a packet number, a total packet number and cut data of the compressed three-dimensional sonar data.

6. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 5, wherein, In the second deep learning network, the second MLP layer, the second Attention layer and the second Point Transformer layer are used to respectively restore the aggregated information of the deleted points in the compressed three-dimensional sonar data, the restored data is fused by a third MLP layer to obtain decompressed deleted points, and the reserved points and the decompressed deleted points are subjected to reflection intensity data recovery based on the maximum reflection intensity of the current frame and the minimum reflection intensity of the current frame, and finally, decompressed three-dimensional sonar data is obtained.

7. The deep learning based multi-beam forward-looking three-dimensional sonar data compression system of claim 4, wherein, The application comprises the following steps:

8. A deep learning based multi-beam forward-looking three-dimensional sonar data compression method, implemented by using the deep learning based multi-beam forward-looking three-dimensional sonar data compression system of any one of claims 1-7, characterized in that, ​ The acquired three-dimensional sonar data is normalized in reflection intensity and blocked in point cloud by using a three-dimensional sonar data compression lower computer platform, and is down-sampled by a first deep learning network including a first MLP layer, a first Attention layer and a first Point Transformer layer to obtain compressed three-dimensional sonar data and transmit the compressed three-dimensional sonar data to a three-dimensional sonar data decompression upper computer platform; The received compressed three-dimensional sonar data is up-sampled by a second deep learning network including a second MLP layer, a second Attention layer and a second Point Transformer layer for intensity data recovery by using the three-dimensional sonar data decompression upper computer platform to obtain decompressed three-dimensional sonar data.

9. A deep learning based multi-beam forward-looking three-dimensional sonar data compression device, comprising a memory and one or more processors, the memory being used to store a computer program, characterized in that, The processor is configured to implement the deep learning-based multi-beam forward-looking three-dimensional sonar data compression method of claim 8 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is configured to implement the deep learning-based multi-beam forward-looking three-dimensional sonar data compression method of claim 8 when executed by a computer.

Images