A system for optimizing data compression using neural networks

Abstract

The application is a data compression system optimized by neural network in the field of data compression, aiming at solving the problems of large storage redundancy and low restoration accuracy of original trajectory data. The process is as follows: first, the sensor collects the pen tip motion trajectory coordinate points, and divides the independent sub-trajectory segments according to the speed change, direction turning and writing pause characteristics; then, equal distance resampling and normalization processing are performed on each sub-trajectory segment to generate the first data containing normalized coordinate increment sequence and auxiliary metadata; then, a four-layer fully connected neural network encoder is used to compress high-dimensional coordinate increments into low-dimensional hidden variables and quantitatively store them, and the auxiliary metadata is combined to generate a compressed file; when decompressing, the hidden variables are dequantized and restored, the coordinate increments are generated through the decoder, and the complete trajectory is obtained by combining the auxiliary metadata and accumulating. Through the design of the loss function balancing the comprehensive point position offset, length difference and direction difference, it is suitable for the storage and transmission of whiteboard handwriting data.

Description

Technical Field

[0001] This invention relates to the field of data compression, specifically a system that uses neural networks to optimize data compression. Background Technology

[0002] With the popularization of digital teaching and remote collaboration, the demand for storage and transmission of handwritten trajectory data generated by whiteboards, as an efficient handwriting interaction tool, has increased significantly. Whiteboard handwritten trajectories consist of a large number of continuously sampled coordinate points, including key geometric features such as point accuracy, length, and direction. Moreover, the characteristics of different trajectories vary greatly, which poses a dual requirement for data compression technology: efficient compression and feature preservation.

[0003] Existing trajectory data compression technologies have significant limitations: general compression algorithms are not optimized for trajectory geometry, easily leading to trajectory distortion after decompression; simple downsampling reduces data volume but loses crucial details; and the lack of adaptive processing for different trajectory types makes it difficult to balance compression ratio and restoration accuracy. These problems result in large data redundancy and low transmission efficiency in trajectory data storage, and geometric feature distortion is likely to occur after decompression, restricting the application of whiteboard data in scenarios such as real-time interaction and long-term storage.

[0004] Based on this, the present invention proposes a method for compressing and decompressing whiteboard handwritten trajectory data, aiming to solve the above pain points and achieve a balance between efficient compression and accurate restoration through targeted segmentation, preprocessing and intelligent compression strategies. Summary of the Invention

[0005] To address the technical problems mentioned in the background art, the present invention adopts the following technical solution: a system for data compression optimized using a neural network, the system comprising:

[0006] Data acquisition and processing module: The module acquires the coordinates of the continuous movement trajectory of the pen tip on the whiteboard through sensors, divides the trajectory into initial independent sub-trajectory segments, and then preprocesses the initial independent sub-trajectory segments to obtain the first data.

[0007] Data compression and decompression module: This module is used to input the first data into the autoencoder model, use the encoder and decoder of the autoencoder model to perform data compression and decompression operations, and finally output the decompressed trajectory coordinates and compression ratio.

[0008] Furthermore, the data acquisition process is as follows:

[0009] The coordinates of the pen tip's movement trajectory are continuously sampled by a sensor, and the coordinates of the pen tip's movement trajectory constitute a set of discrete points.

[0010] Based on the characteristics of velocity changes, direction changes, and writing pauses at the coordinate points of the motion trajectory, the trajectory segment positions are determined, and the trajectory is divided into initial independent sub-trajectory segments, represented as:

[0011]

[0012] Where K represents the total number of initial independent sub-trajectory segments in the pen tip movement trajectory division; The m-th initial independent sub-trajectory segment is represented as:

[0013]

[0014] in, Represents the m-th initial independent sub-trajectory segment. j The coordinates of the pen tip's movement trajectory. N This represents the total number of trajectory coordinate points contained in the m-th initial independent sub-trajectory segment.

[0015] Furthermore, the data preprocessing process is as follows:

[0016] Each of the initial independent sub-trajectory segments is preprocessed to obtain the final independent sub-trajectory segments. The data preprocessing includes resampling and normalization.

[0017] The resampling involves performing equidistant linear interpolation resampling on each of the initial independent sub-trajectory segments, as follows:

[0018] The first step is to calculate the actual length of the initial independent sub-trajectory segment. The formula for calculating the actual length is:

[0019]

[0020] The second step is to set the target sampling interval based on the trajectory complexity. The trajectory complexity is represented by the average rate of change of the slope of the line connecting adjacent points of the initial independent sub-trajectory segment;

[0021] Initial independent sub-trajectory segments The average rate of change of the slope of the line connecting adjacent points is calculated using the following formula:

[0022]

[0023] in, Indicates the initial independent sub-trajectory segment The rate of change of the slope of the line connecting adjacent points is calculated using the following formula:

[0024]

[0025] in: Indicates the initial independent sub-trajectory segment adjacent coordinate points and The slope of the line connecting the two sides is calculated using the following formula:

[0026]

[0027] The target sampling interval The settings should follow these rules:

[0028] Rule 1: If The trajectory was determined to be highly complex, and the target sampling interval was set to be... ;

[0029] Rule 2: If The trajectory was determined to be of low complexity, and the target sampling interval was set to be... ;

[0030] in, This indicates a preset slope change rate threshold, which is determined manually.

[0031] The initial independent sub-trajectory segment The sampling interval is The calculation formula is:

[0032]

[0033] The third step is to perform equidistant linear interpolation on the initial independent sub-trajectory segments, uniformly generating them over the actual length. A new sampling point is used to replace the coordinates of the initial independent sub-trajectory segment with the coordinates of the new sampling point, resulting in the final independent sub-trajectory segment. The corresponding final independent sub-trajectory segment is represented as follows:

[0034]

[0035] Wherein, sub-coordinates For the final independent sub-trajectory segment The starting point, sub-coordinates For the final independent sub-trajectory segment end;

[0036] For the final independent sub-trajectory segment The normalization process is as follows:

[0037] First, calculate the final independent sub-trajectory segments. adjacent coordinate increments The calculation formula is:

[0038]

[0039] in ;

[0040] Using the actual length The final independent sub-trajectory segment The coordinate increments are normalized to the (-1, 1) interval:

[0041]

[0042] in, For the final independent sub-trajectory segment The normalized coordinate increment of the j-th sampling interval;

[0043] The preprocessed data is used as the first data, which includes a normalized coordinate increment sequence and auxiliary metadata.

[0044] The final independent sub-trajectory segment in the first data The corresponding normalized coordinate increment sequence is expressed as:

[0045]

[0046] in, This represents the normalized coordinate increment sequence of the m-th final independent sub-trajectory segment, with dimension . ;

[0047] The final independent sub-trajectory segment in the first data The corresponding auxiliary metadata is represented as follows:

[0048] .

[0049] Furthermore, the process of data compression using an encoder is as follows:

[0050] Step 2-1: Using the first data as input, the normalized coordinate increment sequence in the first data is used to form an input set, which is then compressed. Represented as:

[0051]

[0052] Step 2-2: Data compression is performed using a four-layer fully connected neural network encoder. The process is as follows:

[0053] First, the normalized coordinate increment sequence is processed. Perform forward propagation to obtain the corresponding low-dimensional hidden variables. The low-dimensional latent variables The formula is:

[0054] ;

[0055] Indicates the encoding function;

[0056] Then, the low-dimensional latent variables... Quantization is performed to obtain the quantized low-dimensional latent variables. , The calculation formula is:

[0057]

[0058] Where: B represents the quantization bit depth, which is determined by observing the autoencoder model under different... B The training results under the given value are used to determine the value; Represents a mapping function;

[0059] Steps 2-3: Quantify the low-dimensional latent variables Auxiliary metadata in the first data Merge and generate a compressed file , is represented as:

[0060]

[0061] Repeat step 2-2 to compress the K final independent sub-trajectory segments, generating K compressed files;

[0062] Steps 2-4: Calculate the final independent sub-trajectory segments The compression ratio R is given by the following formula:

[0063]

[0064] in, Indicates the final independent sub-trajectory segment Compressed data dimensions Indicates the final independent sub-trajectory segment The dimension corresponding to the normalized coordinate increment sequence;

[0065] Steps 2-5: Store the K compressed files using storage units, employing a binary serialization format.

[0066] Furthermore, the process of data decompression using the decoder of the autoencoder model is as follows:

[0067] Step 3-1: Using the compressed files stored in the storage unit as input, combine the K compressed files into a compressed file set. The Represented as:

[0068]

[0069] Step 3-2: Traverse the set of compressed files Read each compressed file and verify its integrity. The verification includes:

[0070] Verification 1: Check the low-dimensional latent variables after quantization Is the dimension of ;

[0071] Verification 2: Verify auxiliary metadata Is the format correct?

[0072] Step 3-3: Decompress the compressed file that has passed verification. The specific process is as follows:

[0073] First, regarding the first o Quantized low-dimensional latent variables in a compressed file Perform dequantization to obtain the low-dimensional latent variables before quantization. The calculation formula is:

[0074]

[0075] Then, for the inverse quantization Perform forward propagation and output the restored normalized coordinate increments, using the following formula:

[0076]

[0077] in, , Indicates the decoding function; Indicates the starting point of the normalized coordinate increment after restoration; This indicates the endpoint of the normalized coordinate increment after restoration;

[0078] Secondly, in conjunction with the first o Auxiliary metadata in a compressed file In Regarding the above Perform inverse normalization and output the restored coordinates. The calculation formula is as follows:

[0079]

[0080] in, ; Indicates the first o Length of each initial independent sub-trajectory segment; express The second one in j One element; express The second one in j +1 element;

[0081] Finally, in conjunction with the first o Auxiliary metadata in a compressed file The starting coordinates in The coordinates of the decompressed trajectory are obtained by summing the results.

[0082]

[0083] Steps 3-4: Output the following data:

[0084] The trajectory coordinates after decompression;

[0085] The original whiteboard drawings or handwritten notes used for restoration;

[0086] Compression ratio R.

[0087] Compared with the prior art, the advantages of the present invention are as follows:

[0088] 1. This invention divides the trajectory into independent sub-trajectory segments based on speed changes, direction changes, and writing pauses. By combining equidistant resampling and normalization processing, it generates first data containing a normalized coordinate increment sequence and auxiliary metadata. This solves the problems of traditional compression techniques not addressing trajectory geometric features and insufficient feature extraction of sub-trajectory segments, laying the foundation for efficient compression.

[0089] 2. This invention employs a four-layer fully connected neural network encoder to compress high-dimensional coordinate increments into low-dimensional latent variables and quantize and store them. Combined with auxiliary metadata, a compressed file is generated. Trajectory restoration is achieved through dequantization, decoding, and denormalization, which solves the problems of low compression efficiency and poor restoration accuracy of traditional compression algorithms, while balancing compression ratio and trajectory integrity.

[0090] 3. This invention designs a loss function that integrates point offset, length difference, and direction difference. By using weighting coefficients to balance the impact of different errors on the compression effect, it solves the problem that traditional loss functions only focus on a single error and cannot guarantee the consistency of trajectory geometric features, thereby improving the quality of trajectory reconstruction after decompression. Attached Figure Description

[0091] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0092] Figure 1 This is a schematic diagram of the system flow of the present invention;

[0093] Figure 2This is a flowchart of the data compression and decompression module of the present invention. Detailed Implementation

[0094] To achieve the above objectives, the present invention provides a system for optimizing data compression using a neural network. A schematic diagram of the system flow is provided below. Figure 1 The specific implementation process includes:

[0095] Step 1: Data Collection

[0096] When writing notes or drawing on a whiteboard, sensors continuously sample the coordinates of the pen tip's movement trajectory, and these coordinates form a set of discrete points. Where i = 1, 2, ..., T, and T is the total number of sampling points for the pen tip motion trajectory;

[0097] Based on the characteristics of velocity changes, directional changes, and writing pauses at the coordinate points of the pen tip's movement trajectory, the trajectory segmentation positions are determined, and the trajectory is divided into initial independent sub-trajectory segments, with the following specific rules:

[0098] The sampling points t speed The calculation formula is:

[0099]

[0100] in, The sampling time interval is determined by the sensor frequency; Represents the trajectory coordinates of sampling point t;

[0101] If the velocities at k consecutive points satisfy This is identified as a writing pause, triggering a segmentation.

[0102] in, Indicates the speed threshold, the The determination is based on statistical learning of users' historical writing data, continuously updating historical data and adjusting in real time. To adapt to user habits;

[0103] Direction angle change of adjacent sampling points The calculation formula is:

[0104]

[0105] in, Indicates the magnitude of the direction angle at sampling point t;

[0106] like The system is determined to have experienced a sharp change in direction, triggering a segmentation process.

[0107] in, Indicates the direction angle threshold, the Determined based on statistical analysis of the directional angles recorded in historical texts;

[0108] The initial independent sub-trajectory segment is represented as:

[0109]

[0110] Where K represents the total number of initial independent sub-trajectory segments in the pen tip movement trajectory division; The m-th initial independent sub-trajectory segment is represented as:

[0111]

[0112] in, Represents the m-th initial independent sub-trajectory segment. j The coordinates of the pen tip's movement trajectory. N This represents the total number of trajectory coordinate points contained in the m-th initial independent sub-trajectory segment.

[0113] Step 2: Data Preprocessing

[0114] Each of the initial independent sub-trajectory segments is preprocessed to obtain the final independent sub-trajectory segments. The data preprocessing includes resampling and normalization.

[0115] The resampling involves performing equidistant linear interpolation resampling on each of the initial independent sub-trajectory segments, as follows:

[0116] The first step is to calculate the actual length of the initial independent sub-trajectory segment. The formula for calculating the actual length is:

[0117]

[0118] The second step is to set the target sampling interval based on the trajectory complexity. The trajectory complexity is represented by the average rate of change of the slope of the line connecting adjacent points of the initial independent sub-trajectory segment;

[0119] Initial independent sub-trajectory segments The average rate of change of the slope of the line connecting adjacent points is calculated using the following formula:

[0120]

[0121] in, Indicates the initial independent sub-trajectory segment The rate of change of the slope of the line connecting adjacent points is calculated using the following formula:

[0122]

[0123] in: Indicates the initial independent sub-trajectory segment adjacent coordinate points and The slope of the line connecting the two sides is calculated using the following formula:

[0124]

[0125] The target sampling interval The settings should follow these rules:

[0126] Rule 1: If The trajectory was determined to be highly complex, and the target sampling interval was set to be... ;

[0127] Rule 2: If The trajectory was determined to be of low complexity, and the target sampling interval was set to be... ;

[0128] in, This indicates a preset slope change rate threshold, which is determined manually.

[0129] The initial independent sub-trajectory segment The sampling interval is The calculation formula is:

[0130]

[0131] The third step is to perform equidistant linear interpolation on the initial independent sub-trajectory segments, uniformly generating them over the actual length. A new sampling point is used to replace the coordinates of the initial independent sub-trajectory segment with the coordinates of the new sampling point, resulting in the final independent sub-trajectory segment. The corresponding final independent sub-trajectory segment is represented as follows:

[0132]

[0133] Wherein, sub-coordinates For the final independent sub-trajectory segment The starting point, sub-coordinates For the final independent sub-trajectory segment end;

[0134] For the final independent sub-trajectory segment The normalization process is as follows:

[0135] First, calculate the final independent sub-trajectory segments. adjacent coordinate increments The calculation formula is:

[0136]

[0137] in ;

[0138] Using the actual length The final independent sub-trajectory segment The coordinate increments are normalized to the (-1, 1) interval:

[0139]

[0140] in, For the final independent sub-trajectory segment The normalized coordinate increment of the j-th sampling interval;

[0141] The preprocessed data is used as the first data, which includes a normalized coordinate increment sequence and auxiliary metadata.

[0142] The final independent sub-trajectory segment in the first data The corresponding normalized coordinate increment sequence is expressed as:

[0143]

[0144] in, This represents the normalized coordinate increment sequence of the m-th final independent sub-trajectory segment, with dimension . ;

[0145] The final independent sub-trajectory segment in the first data The corresponding auxiliary metadata is represented as follows:

[0146] .

[0147] Step 3: Data Compression and Decompression

[0148] Data compression and decompression are performed using an autoencoder model constructed from an autoencoder neural network, as shown in the flowchart below. Figure 2 As shown, the specific steps are as follows:

[0149] (1) Constructing an autoencoder model

[0150] An autoencoder model containing an encoder and a decoder is constructed using an autoencoder neural network to achieve compression and restoration of high-dimensional coordinate incremental data;

[0151] The encoder encodes high-dimensional input data into low-dimensional latent variables.

[0152] The decoder: decodes the encoded latent variables into high-dimensional output data, restoring the coordinate increments;

[0153] Construct a four-layer fully connected neural network autoencoder model, wherein the four-layer fully connected neural network includes:

[0154] Input layer: The number of neurons is consistent with the dimension of the normalized coordinate increment sequence of the input;

[0155] First hidden layer: The number of neurons is the same as the number of neurons in the input layer. ReLU is used as the activation function;

[0156] Second hidden layer: The number of neurons is the same as the number of neurons in the first hidden layer. ReLU is used as the activation function;

[0157] Output layer: Tanh is used as the activation function;

[0158] Loss function design:

[0159] The loss function is constructed by minimizing point offset, length difference, and orientation difference. The expression for the loss function is:

[0160]

[0161] in, , and The weighting coefficients are dynamically adjusted based on the complexity of the trajectory segments. The core logic is as follows:

[0162] If the trajectory complexity is high, for It is more sensitive;

[0163] If the trajectory complexity is low, for and It is more sensitive;

[0164] LOC represents the point offset error, which measures the spatial deviation between the normalized coordinate increment and the decoded coordinate increment in the first data. The formula is:

[0165]

[0166] in, For the m-th final independent sub-trajectory segment, the normalized coordinate increment in the first data is... The coordinate increment of the decoded output of the m-th final independent sub-trajectory segment; K represents the total number of initial independent sub-trajectory segments into which the pen tip motion trajectory is divided;

[0167] LEN stands for Length Difference Error, which measures the length deviation between the first data and the decoded trajectory. The calculation formula is:

[0168]

[0169] LDM stands for Orientation Difference Error, which measures the overall orientation deviation between the first data and the decoded trajectory. The calculation formula is:

[0170]

[0171] in, These are the azimuth angles of the first data and the decoded coordinate increment of the m-th initial independent sub-trajectory segment, respectively;

[0172] (2) Data compression

[0173] The steps for data compression using an encoder are as follows:

[0174] First, the first data is used as input. The normalized coordinate increment sequence in the first data is used to form an input set, which is then compressed. Represented as:

[0175]

[0176] Secondly, a four-layer fully connected neural network encoder is used for data compression, the process of which is as follows:

[0177] First, the normalized coordinate increment sequence is processed. Perform forward propagation to obtain the corresponding low-dimensional hidden variables. The low-dimensional latent variables The formula is:

[0178] ;

[0179] Indicates the encoding function;

[0180] Then, the low-dimensional latent variables... Quantization is performed to obtain the quantized low-dimensional latent variables. , The calculation formula is:

[0181]

[0182] Where: B represents the quantization bit depth, which is determined by observing the autoencoder model under different... B The training results under the given value are used to determine the value; Represents a mapping function;

[0183] quantized low-dimensional latent variables Auxiliary metadata in the first data Merge and generate a compressed file , is represented as:

[0184]

[0185] Compress the K final independent sub-trajectory segments to generate K compressed files;

[0186] Calculate the final independent sub-trajectory segments The compression ratio R is given by the following formula:

[0187]

[0188] in, Indicates the final independent sub-trajectory segment Compressed data dimensions Indicates the final independent sub-trajectory segment The dimension corresponding to the normalized coordinate increment sequence;

[0189] The K compressed files are stored using storage units, and the storage method adopts binary serialization format.

[0190] (3) Data decompression

[0191] Step 3-1: Using the compressed files stored in the storage unit as input, combine the K compressed files into a compressed file set. The Represented as:

[0192]

[0193] Step 3-2: Traverse the set of compressed files Read each compressed file and verify its integrity. The verification includes:

[0194] Verification 1: Check the low-dimensional latent variables after quantization Is the dimension of ;

[0195] Verification 2: Verify auxiliary metadata Is the format correct?

[0196] Step 3-3: Decompress the compressed file that has passed verification. The specific process is as follows:

[0197] First, regarding the first o Quantized low-dimensional latent variables in a compressed file Perform dequantization to obtain the low-dimensional latent variables before quantization. The calculation formula is:

[0198]

[0199] Then, for the inverse quantization Perform forward propagation and output the restored normalized coordinate increments, using the following formula:

[0200]

[0201] in, , Indicates the decoding function; Indicates the starting point of the normalized coordinate increment after restoration; This indicates the endpoint of the normalized coordinate increment after restoration;

[0202] Secondly, in conjunction with the first o Auxiliary metadata in a compressed file In Regarding the above Perform inverse normalization and output the restored coordinates. The calculation formula is as follows:

[0203]

[0204] in, ; Indicates the first o Length of each initial independent sub-trajectory segment; express The second one in j One element; express The second one in j +1 element;

[0205] Finally, in conjunction with the first o Auxiliary metadata in a compressed file The starting coordinates in The coordinates of the decompressed trajectory are obtained by summing the results.

[0206]

[0207] Steps 3-4: Output the following data:

[0208] The trajectory coordinates after decompression;

[0209] The original whiteboard drawings or handwritten notes used for restoration;

[0210] Compression ratio R.

[0211] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for optimizing data compression using neural networks, characterized in that, The system includes: Data acquisition and processing module: The module acquires the coordinates of the continuous movement trajectory of the pen tip on the whiteboard through sensors, divides the trajectory into initial independent sub-trajectory segments, and then preprocesses the initial independent sub-trajectory segments to obtain the first data. Data compression and decompression module: used to input the first data into the autoencoder model, use the encoder and decoder of the autoencoder model to perform data compression and decompression operations, and finally output the decompressed trajectory coordinates and compression ratio; The data acquisition process is as follows: The coordinates of the pen tip's movement trajectory are continuously sampled by a sensor, and the coordinates of the pen tip's movement trajectory constitute a set of discrete points. Based on the characteristics of velocity changes, direction changes, and writing pauses at the coordinate points of the motion trajectory, the trajectory segment positions are determined, and the trajectory is divided into initial independent sub-trajectory segments, represented as: Where K represents the total number of initial independent sub-trajectory segments in the pen tip movement trajectory division; The m-th initial independent sub-trajectory segment is represented as: in, and This represents the coordinates of the j-th pen tip movement trajectory in the m-th initial independent sub-trajectory segment, and N represents the total number of trajectory coordinates contained in the m-th initial independent sub-trajectory segment. The data preprocessing process is as follows: Each of the initial independent sub-trajectory segments is preprocessed to obtain the final independent sub-trajectory segments. The data preprocessing includes resampling and normalization. The resampling involves performing equidistant linear interpolation resampling on each of the initial independent sub-trajectory segments, as follows: The first step is to calculate the actual length of the initial independent sub-trajectory segment. The formula for calculating the actual length is: The second step is to set the target sampling interval based on the trajectory complexity. The trajectory complexity is represented by the average rate of change of the slope of the line connecting adjacent points of the initial independent sub-trajectory segment; Initial independent sub-trajectory segments The average rate of change of the slope of the line connecting adjacent points is calculated using the following formula: in, Indicates the initial independent sub-trajectory segment The rate of change of the slope of the line connecting adjacent points is calculated using the following formula: in: Indicates the initial independent sub-trajectory segment adjacent coordinate points and The slope of the line connecting the two sides is calculated using the following formula: The target sampling interval The settings should follow these rules: Rule 1: If The trajectory was determined to be highly complex, and the target sampling interval was set to be... ; Rule 2: If The trajectory was determined to be of low complexity, and the target sampling interval was set to be... ; in, This indicates the preset slope change rate threshold, which is determined manually.

2. The system according to claim 1, characterized in that, The initial independent sub-trajectory segment The sampling interval is The calculation formula is: The third step is to perform equidistant linear interpolation on the initial independent sub-trajectory segments, uniformly generating them over the actual length. A new sampling point is used to replace the coordinates of the initial independent sub-trajectory segment with the coordinates of the new sampling point, resulting in the final independent sub-trajectory segment. The corresponding final independent sub-trajectory segment is represented as follows: Wherein, sub-coordinates For the final independent sub-trajectory segment The starting point, sub-coordinates For the final independent sub-trajectory segment end; For the final independent sub-trajectory segment The normalization process is as follows: First, calculate the final independent sub-trajectory segments. adjacent coordinate increments and The calculation formula is: in ; Using the actual length The final independent sub-trajectory segment The coordinate increments are normalized to the (-1, 1) interval: in, and For the final independent sub-trajectory segment The normalized coordinate increment of the j-th sampling interval; The preprocessed data is used as the first data, which includes a normalized coordinate increment sequence and auxiliary metadata. The final independent sub-trajectory segment in the first data The corresponding normalized coordinate increment sequence is expressed as: in, Indicates the first The normalized coordinate increment sequence of each final independent sub-trajectory segment, with dimension [ ]. ; The final independent sub-trajectory segment in the first data The corresponding auxiliary metadata is represented as follows: 。 3. The system according to claim 2, characterized in that, The process of data compression using an encoder is as follows: Step 2-1: Using the first data as input, the normalized coordinate increment sequence in the first data is used to form an input set, which is then compressed. Represented as: Step 2-2: Data compression is performed using a four-layer fully connected neural network encoder. The process is as follows: First, the normalized coordinate increment sequence is processed. Perform forward propagation to obtain the corresponding low-dimensional hidden variables. The low-dimensional latent variables The formula is: ； Indicates the encoding function; Then, the low-dimensional latent variables... Quantization is performed to obtain the quantized low-dimensional latent variables. , The calculation formula is: in: The quantization bit depth is indicated by observing the autoencoder model at different... B The training results under the given value are used to determine the value; Represents a mapping function; Steps 2-3: Quantify the low-dimensional latent variables Auxiliary metadata in the first data Merge and generate a compressed file , is represented as: Repeat step 2-2 to compress the K final independent sub-trajectory segments, generating K compressed files; Steps 2-4: Calculate the final independent sub-trajectory segments The compression ratio R is given by the following formula: in, Indicates the final independent sub-trajectory segment Compressed data dimensions Indicates the final independent sub-trajectory segment The dimension corresponding to the normalized coordinate increment sequence; Steps 2-5: Store the K compressed files using storage units, employing a binary serialization format.

4. The system according to claim 3, characterized in that, The process of data decompression using the decoder of the autoencoder model is as follows: Step 3-1: Using the compressed files stored in the storage unit as input, combine the K compressed files into a compressed file set. The Represented as: Step 3-2: Traverse the set of compressed files Read each compressed file and verify its integrity. The verification includes: Verification 1: Check the low-dimensional latent variables after quantization Is the dimension of ; Verification 2: Verify auxiliary metadata Is the format correct? Step 3-3: Decompress the compressed file that has passed verification. The specific process is as follows: First, the quantized low-dimensional latent variables in the o-th compressed file are... Perform dequantization to obtain the low-dimensional latent variables before quantization. The calculation formula is: Then, for the inverse quantization Perform forward propagation and output the restored normalized coordinate increments, using the following formula: in, , Indicates the decoding function; and Indicates the starting point of the normalized coordinate increment after restoration; and This indicates the endpoint of the normalized coordinate increment after restoration; Secondly, combining the auxiliary metadata in the o-th compressed file In Regarding the above Perform inverse normalization and output the restored coordinates. The calculation formula is as follows: in, ; This represents the length of the 0th initial independent sub-trajectory segment; express The first in 2j One element; express The first in 2j+1 One element; Finally, combining the auxiliary metadata in the o-th compressed file The starting coordinates in The coordinates of the decompressed trajectory are obtained by summing the results. Steps 3-4: Output the following data: The trajectory coordinates after decompression; The original whiteboard drawings or handwritten notes used for restoration; Compression ratio R.

Images