Feature data transmission method, device and equipment and computer readable storage medium

Abstract

The invention discloses a feature data transmission method, device and equipment and a computer readable storage medium, and relates to the technical field of data transmission, initial feature data to be transmitted is acquired, and amplitude scalar data is determined according to the initial feature data; performing amplitude quantization according to the amplitude scalar data to obtain amplitude quantization data, and performing directional quantization according to the initial feature data and the amplitude scalar data to obtain directional quantization data; and determining target transmission data according to the amplitude quantized data and the direction quantized data so as to realize feature data transmission. According to the invention, the accuracy of feature data transmission is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data transmission, in particular to a feature data transmission method and device, equipment and a computer readable storage medium. BACKGROUND

[0002] With the development of data transmission, especially feature data transmission, users also have higher requirements for feature data transmission methods.

[0003] The traditional feature data transmission method is to compress the data after using an implicit codebook, that is, to independently map each feature dimension to a few discrete scalars, and all dimensions are combined to form an implicit code word space. This feature data transmission method has certain defects, which may cause the phenomenon that each dimension is quantized independently in the compression process, and each dimension is equal in weight and independent of each other (that is, when the feature components are greatly different or have correlation, the transmitted data may have errors, causing errors in subsequent recovery). That is, this feature data transmission method may cause low accuracy of feature data transmission due to the independent quantization of each dimension in the compression process and the default equal weight and independence of each dimension.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a feature data transmission method, device, equipment and computer readable storage medium, which aims to solve the technical problem of low accuracy of feature data transmission.

[0006] To achieve the above purpose, the present application provides a feature data transmission method, which comprises: obtaining initial feature data to be transmitted, and determining amplitude scalar data according to the initial feature data; According to the amplitude scalar data, the amplitude quantization data is obtained by amplitude quantization, and the direction quantization data is obtained by direction quantization according to the initial feature data and the amplitude scalar data; According to the amplitude quantization data and the direction quantization data, the target transmission data is determined to realize feature data transmission.

[0007] In an embodiment, the step of amplitude quantization according to the amplitude scalar data to obtain amplitude quantization data comprises: Based on the preset amplitude quantization function, the amplitude scalar data is amplitude quantized to obtain a plurality of first quantization data, and the amplitude target interval of each first quantization data in the preset quantization range interval is determined; determining first encoding data corresponding to each of the amplitude target intervals, and taking the first encoding data as amplitude quantization data.

[0008] In an embodiment, the step of determining first encoding data corresponding to each of the amplitude target intervals comprises: determining the number of intervals of the amplitude target intervals, and encoding each of the amplitude target intervals based on target encoding corresponding to the number of intervals to obtain first encoding data.

[0009] In an embodiment, the step of obtaining direction quantization data according to the initial feature data and the amplitude scalar data comprises: determining a ratio between the initial feature data and the amplitude scalar data as unit direction data, and performing direction quantization on the unit direction data based on a preset direction quantization function to obtain a plurality of second quantization data, and determining a direction target interval of each of the second quantization data in a preset quantization range interval. determining second encoding data corresponding to each of the direction target intervals, and taking the second encoding data as direction quantization data.

[0010] In an embodiment, the feature data transmission method further comprises: obtaining a distribution range of quantization data, wherein the quantization data comprises first quantization data about amplitude and second quantization data about direction; performing uniform division based on the distribution range of the quantization data to obtain a preset quantization range interval, or performing division based on the distribution range of the quantization data to obtain a preset quantization range interval according to distribution data amount, wherein the distribution data amount comprises distribution amount of quantization data in different range intervals.

[0011] In an embodiment, after the step of obtaining initial feature data to be transmitted, the feature data transmission method further comprises: obtaining a current transmission environment at a current time, wherein the current transmission environment comprises real-time network state and transmission task demand; in a case where the real-time network state and the transmission task demand both satisfy a preset transmission condition, performing dimension reduction transformation on the initial feature data with a preset first dimension reduction parameter, and taking a dimension-reduced feature vector after the dimension reduction transformation as initial feature data.

[0012] In an embodiment, the feature data transmission method further comprises: If the real-time network status or the transmission task requirements do not meet the preset transmission conditions, the initial feature data is transformed by a preset second dimensionality reduction parameter, and the dimensionality-reduced feature vector after the dimensionality reduction transformation is used as the initial feature data, wherein the target dimension of the second dimensionality reduction parameter is smaller than the target dimension of the first dimensionality reduction parameter.

[0013] Furthermore, to achieve the above objectives, this application also provides a feature data transmission device, the feature data transmission device comprising: The data acquisition module is used to acquire the initial feature data to be transmitted and determine the amplitude scalar data based on the initial feature data. The data quantization module is used to perform amplitude quantization based on the amplitude scalar data to obtain amplitude quantized data, and to perform direction quantization based on the initial feature data and the amplitude scalar data to obtain direction quantized data; The data transmission module is used to determine the target transmission data based on the amplitude quantization data and the direction quantization data, so as to realize the characteristic data transmission.

[0014] In addition, to achieve the above objectives, this application also provides a feature data transmission device, including a processor, a memory, and a feature data transmission method program stored in the memory that can be executed by the processor, wherein when the feature data transmission method program is executed by the processor, it implements the steps of the feature data transmission method as described above.

[0015] This application also provides a computer-readable storage medium storing a feature data transmission method program, wherein when the feature data transmission method program is executed by a processor, it implements the steps of the feature data transmission method as described above.

[0016] This application provides a feature data transmission method. It involves acquiring initial feature data to be transmitted, determining amplitude scalar data based on the initial feature data, performing amplitude quantization on the amplitude scalar data to obtain amplitude quantized data, and performing direction quantization on the initial feature data and amplitude scalar data to obtain direction quantized data. Finally, it determines target transmission data based on the amplitude quantized data and direction quantized data to achieve feature data transmission. This method determines amplitude scalar data based on initial feature data, processes the amplitude scalar data to obtain target transmission data, and thus achieves feature data transmission. Furthermore, it allows direct quantization of the data to be transmitted to achieve data transmission. Specifically, this feature data transmission method directly performs amplitude and direction quantization on the initial feature data to obtain target transmission data, and then performs data transmission based on the target transmission data to achieve data transmission, thereby reducing the cost of feature data transmission. Moreover, the entire process separates the amplitude quantization data and direction quantization data, avoiding independent quantization of each dimension during compression, and assuming that each dimension is equally weighted and independent, thus ensuring the accuracy of feature data transmission. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the feature data transmission method of this application; Figure 2 This is a schematic diagram of a scenario for the feature data transmission method of this application; Figure 3 This is a flowchart illustrating the feature data transmission method of this application; Figure 4 This is a schematic diagram of the modules of the system base chip in this application; Figure 5 This is a schematic diagram of the hardware operating environment involved in the device in this application.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings.

[0019] Explanation of icon numbers: 1001. Processing device; 1002. Read-only memory; 1003. Storage device; 1004. Random access memory; 1005. Bus; 1006. Input / output interface; 1007. Input device; 1008. Output device; 1009. Communication device. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Currently, feature data transmission methods generally employ explicit and implicit codebook schemes. Explicit codebooks can utilize vector quantization and its variant, VQ-VAE (Vector Quantized Variational Autoencoder). This involves obtaining a representative set of codewords through offline training and approximating each input feature to its nearest-neighbor codeword index for compression. While these methods offer high compression ratios, they require storing the codebook simultaneously at the edge and in the cloud, and the training process is complex, often resulting in codebook collapse (where most codewords are not effectively used), leading to high costs associated with feature data transmission. Implicit codebook schemes, on the other hand, do not rely on pre-stored codeword tables. For example, FSQ (Finite Scalar Quantization) maps each feature dimension independently to a few discrete scalars, with all dimensions combined to form an implicit codeword space. Because FSQ only requires simple arithmetic operations such as rounding to quantize each dimension without the need for an additional codebook, it offers more stable training and simpler implementation. However, FSQ quantizes each dimension independently, assuming equal weight and independence for each dimension. This may not be the optimal compression method when feature components differ significantly or are correlated. For example, when the overall amplitude of a visual feature vector varies greatly, FSQ cannot effectively utilize cross-dimensional amplitude information, potentially leading to large quantization errors in some dimensions and increased overall distortion. Another implicit method, BSQ (Binary Spherical Quantization), ensures that the upper limit of quantization error is controlled by normalizing the vector's norm before binarization. BSQ has achieved good results in many scenarios; however, it compresses the amplitude information of all feature vectors to the unit norm, which may result in the loss of important information when amplitude differences are significant, necessitating additional mechanisms (such as contrast alignment correction) to compensate. Based on the above analysis of implicit codebooks, it is found that existing implicit codebook quantization methods suffer from the above defects, resulting in low data transmission accuracy.

[0023] Therefore, based on the shortcomings of the above-mentioned feature data transmission methods, the feature data transmission method of this application is proposed. The solution of the embodiment of this application is: by determining the amplitude scalar data according to the initial feature data, and then processing the amplitude scalar data to obtain the target transmission data, feature data transmission is realized. Furthermore, the data to be transmitted can be directly quantized to achieve data transmission. In other words, this feature data transmission method can directly quantize the initial feature data to be transmitted in amplitude and direction to obtain the target transmission data, and then perform data transmission based on the target transmission data to achieve data transmission, thereby reducing the cost of feature data transmission. Further, the entire process separates the amplitude quantization data and the direction quantization data, avoiding the independent quantization of each dimension during compression, and assuming that each dimension is equally weighted and mutually independent, thus ensuring the accuracy of feature data transmission.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a device capable of performing the above functions, such as a vehicle control terminal. The following description uses a vehicle control terminal as an example to illustrate this embodiment and the subsequent embodiments.

[0025] Based on this, embodiments of this application provide a feature data transmission method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the feature data transmission method of this application.

[0026] Reference Figure 1 This application provides a feature data transmission method. In a first embodiment of the feature data transmission method, the feature data transmission method includes: Step S10: Obtain the initial feature data to be transmitted, and determine the amplitude scalar data based on the initial feature data; In this embodiment, the feature data transmission method is used in a system for feature data transmission. This system includes at least a transmitting end and a receiving end, such as an edge device and a cloud controller, respectively. The edge device can be a controller or control chip on a vehicle. For example, after the transmitting end receives the data to be transmitted, it processes the data to obtain initial feature data, and then performs amplitude extraction on the initial feature data to obtain amplitude scalar data. For example, refer to... Figure 2 , Figure 2This is a schematic diagram of a scenario for the feature data transmission method of this application. The data to be transmitted is visual multimodal feature data. A visual encoder is used to transmit the tower structure in the visual image back to the cloud as feature data. Of course, other data, including amplitude and orientation features, can also be used. Here, visual multimodal feature data is used as an example. The edge device runs a preset visual model to extract... 3D visual feature vector (For example, the output of a hidden layer in a convolutional neural network) can be used as the data to be transmitted. To reduce the amount of data transmitted, a dimensionality reduction transformation (autoencoder) can be used. Projecting to a lower-dimensional space yields This serves as the initial feature data for subsequent quantization. The initial feature data will then be processed to obtain the amplitude scalar data. For details, please refer to... Figure 3 , Figure 3 This is a flowchart illustrating the feature data transmission method of this application. The amplitude scalar data is obtained through feature extraction and dimensionality reduction. The amplitude scalar data can be obtained by calculating the L2 norm of the initial feature data, or by using other methods to obtain the amplitude scalar data, which is not limited here.

[0027] Step S20: Amplitude quantization is performed based on amplitude scalar data to obtain amplitude quantized data, and direction quantization is performed based on initial feature data and amplitude scalar data to obtain direction quantized data; Step S30: Determine the target transmission data based on the amplitude quantization data and the direction quantization data to realize the transmission of characteristic data.

[0028] In this embodiment, after obtaining the amplitude scalar data to be quantized, amplitude quantization is performed on the amplitude scalar data to obtain amplitude quantized data. Amplitude quantization can transform the amplitude scalar data into multiple discrete data, which are then represented by different intervals for data transmission. The amplitude quantized data refers to the amplitude scalar data after quantization. Simultaneously, direction quantization is performed on the initial feature data and amplitude scalar data to obtain direction quantized data. The direction quantized data refers to the direction scalar data after quantization, which can be obtained based on the initial feature data and amplitude scalar data, for example, the quotient between the amplitude scalar data and the initial feature data. Finally, the amplitude quantized data and direction quantized data are combined to form the target transmission data, which is then transmitted to the receiving end to complete the data transmission. For example, the target transmission data can be a bit-encoded segment, with the first N bits representing the amplitude quantization level index and the last N bits representing the direction quantization index bits. Furthermore, amplitude and direction can be transmitted separately using compression to ensure the accuracy of feature data transmission.

[0029] For example, amplitude quantization data can be obtained by extracting the L2 norm of the initial feature data as amplitude scalar data, and then quantizing this amplitude scalar data into a finite number of levels (e.g., several discrete values ​​using logarithmic or linear scales). The overall size information of each amplitude scalar data is then encoded with a small number of bits to obtain amplitude quantized data, preserving the global scale differences between different features. Simultaneously, dividing the initial feature data by its norm yields a unit direction vector (i.e., direction scalar data), and then performing scalar quantization mapping on each component of this unit vector to obtain direction quantized data. Unlike FSQ, which directly quantizes each dimension of the original vector, this application uses normalization before quantization. Directional quantization still does not require an explicit codebook, but because all vectors are normalized to a uniform scale, the quantization scale can be shared among features, thus utilizing the codeword space more effectively. This avoids the loss of vector magnitude information in pure directional quantization (such as BSQ), enabling a more complete reconstruction of the energy intensity of the feature vector. Meanwhile, the directional components eliminate scale differences due to pre-normalization, and the errors in each dimension are controlled within a uniform range during quantization, which helps maintain semantic consistency before and after quantization, thereby improving the accuracy of feature data transmission. For example, the above feature data transmission method can achieve encoding entirely through simple norm calculation and scalar quantization without introducing any additional model parameters or storage overhead. Feature compression can be completed on edge devices by performing only basic operations such as addition, subtraction, multiplication, division, and comparisons, resulting in minimal computational and storage burden, making it ideal for embedded devices with limited operating environments.

[0030] In one embodiment, comparing the above schemes, this application, compared to FSQ, overcomes the drawback of large quantization errors in FSQ when feature magnitude changes drastically by encoding amplitude separately. This results in smaller global reconstruction errors of compressed features against the original features. Furthermore, due to the unified normalization of directional components, this application avoids the problem of inconsistent quantization scales caused by differences in the dynamic range of each dimension in FSQ, improving the consistency and accuracy of high-dimensional feature quantization. Compared to VQ-VAE, this application eliminates the need for cumbersome codebook training and maintenance, and avoids the risk of codebook collapse. Under similar compression ratios, this method achieves feature reconstruction results comparable to VQ-VAE, but with significantly reduced system complexity. Especially in edge-cloud collaborative applications, it eliminates the burden of storing and searching large codebooks on edge devices, and only performs simple arithmetic operations during runtime, significantly improving system reliability and real-time performance. Compared to BSQ, which uses a global unit norm constraint to control the upper limit of quantization error but completely discards amplitude differences, this application adds an amplitude channel, enabling accurate characterization of features of varying intensities and providing stronger expressive power. Furthermore, at the same compression ratio, this application demonstrates superior performance in terms of reconstruction error and downstream task accuracy, showcasing better overall performance.

[0031] In this embodiment, a feature data transmission method is provided. This method involves acquiring initial feature data to be transmitted, determining amplitude scalar data based on the initial feature data, performing amplitude quantization on the amplitude scalar data to obtain amplitude quantized data, and performing direction quantization on the initial feature data and amplitude scalar data to obtain direction quantized data. Finally, it determines the target transmission data based on the amplitude quantized data and direction quantized data to achieve feature data transmission. This method determines amplitude scalar data based on the initial feature data, processes the amplitude scalar data to obtain the target transmission data, and then directly quantizes the data to be transmitted to achieve data transmission. In other words, this feature data transmission method directly performs amplitude and direction quantization on the initial feature data to be transmitted to obtain the target transmission data, and then performs data transmission based on the target transmission data to achieve data transmission, thus reducing the cost of feature data transmission. Furthermore, the entire process separates the amplitude quantization data and direction quantization data, avoiding independent quantization of each dimension during compression, and assuming that each dimension is equally weighted and independent, thereby ensuring the accuracy of feature data transmission.

[0032] Furthermore, based on the first embodiment of this application applied to a vehicle control terminal described above, a second embodiment of the feature data transmission method of this application is proposed. In this embodiment, step S20, the step of obtaining amplitude quantized data by amplitude quantization based on amplitude scalar data, includes: Step S21: Based on a preset amplitude quantization function, the amplitude scalar data is quantized to obtain multiple first quantized data, and the amplitude target range of each first quantized data in the preset quantization range is determined. Step S22: Determine the first encoded data corresponding to each target range of image values, and use the first encoded data as amplitude quantization data, wherein the first encoded data includes the binary code corresponding to each target range of image values.

[0033] In this embodiment, when calculating the initial feature data After obtaining the magnitude scalar data R from the L2 norm, where, This will convert the amplitude scalar data. Through a preset amplitude quantization function Mapped to multiple first quantized data R q ,For example, This will continue to determine the target amplitude range of each first quantized data within the preset quantization range, such as... Figure 3Amplitude quantization data is obtained through amplitude extraction and quantization. The preset quantization range interval is based on the interval defined by the first quantized data. The preset amplitude quantization function is the function that quantizes the amplitude, such as representing the amplitude linearly or logarithmically. The linear or logarithmic scale is selected according to application requirements. The amplitude target interval refers to the interval corresponding to the first quantized data; for example, the first quantized data 's' corresponds to the amplitude target interval 's1', and 's1' is then represented based on the first encoded data. For example, the possible amplitude range can be divided into several intervals, each interval corresponding to a first quantized data. This allows us to re-determine the first encoded data corresponding to each target value interval, and then use this first encoded data as amplitude quantization data. The first encoded data includes the binary code corresponding to each target value interval. For example, multiple determined first quantization data R... q The intervals are defined as 2, 4, 6, and 8 respectively. Then, the intervals 1, 2, 3, and 4 are defined to represent 2, 4, 6, and 8 respectively. That is, in actual transmission, only the intervals 1, 2, 3, and 4 need to be transmitted. This can be represented using 2-bit binary encoding. For example, 00, 01, 10, and 11 can represent the intervals 1, 2, 3, and 4 respectively. In other words, only 2-bit binary encoding is needed to represent the amplitude quantization data, which can greatly reduce the amount of data transmission.

[0034] Furthermore, the step of determining the first encoded data corresponding to each target range of image values ​​includes: Step S221: Determine the number of amplitude target intervals, and encode each amplitude target interval based on the target code corresponding to the number of intervals to obtain the first coded data.

[0035] In this embodiment, determining the first encoded data can be done by first determining the amount of data to be transmitted, and then determining the number of encoded bits to be used based on the amount of data. Since the transmitted data is generally a continuous range, binary encoding can be used directly. For example, if there are 4 ranges in the amplitude target range, binary encoding can use 2 bits; if there are 8 ranges in the amplitude target range, binary encoding can use 3 bits. For example, when the number of ranges in the amplitude target range is less than the number that the encoding can represent, the encoding can be used, and the excess bits are not used.

[0036] Furthermore, based on the first and / or second embodiments of this application described above, a third embodiment of the feature data transmission method of this application is proposed. In this embodiment, step S20, the step of obtaining direction quantization data by performing direction quantization based on initial feature data and amplitude scalar data, includes: Step S23: Determine the ratio between the initial feature data and the amplitude scalar data as the unit direction data, and perform direction quantization on the unit direction data based on the preset direction quantization function to obtain multiple second quantized data, and determine the direction target interval of each second quantized data in the preset quantization range interval; Step S24: Determine the second encoded data corresponding to the target interval in each direction, and use the second encoded data as the direction quantization data, wherein the second encoded data includes the binary code corresponding to the target interval in each direction.

[0037] In this embodiment, after determining the amplitude scalar data, the unit direction of the eigenvector (unit direction data) is calculated. For example, because of the scalar transformation, the amplitude scalar data is a number greater than or equal to 0, and therefore when... When it is not zero, To and Unit vectors in the same direction with a norm of 1, when Then it can be made Vector. Furthermore, the above steps can be used to decouple the orientation information from the amplitude of the original features, obtaining normalized unit orientation data containing only the orientation. ,like Figure 3 Direction quantization data is obtained through direction normalization and direction quantization encoding. Further, this is applied to unit direction data... Each dimension employs a directional quantization function. Map its components onto a predetermined discrete set. For example, for The Each component Quantize it into multiple second-quantized data. For example, one could choose to use symmetry in each dimension. A set of discrete values ​​(e.g., {-1, -0.5, 0.5, 1} for L=4) is compared by rounding. Mapped to the closest value in the set. After processing all dimensions, multiple second-quantized data points are obtained. Each of these Only a few bits are needed for encoding (depending on the size of the discrete series L, for example, L=4 corresponds to 2 bits, L=8 corresponds to 3 bits, and so on). Based on the amplitude representation method of the above embodiments, the second encoded data corresponding to each directional target interval is determined, and the second encoded data is used as directional quantization data. The second encoded data includes the binary code corresponding to each directional target interval. The division of the preset quantization range interval can also be based on the interval divided by the second quantization data. The preset directional quantization function refers to the function that quantizes the direction, such as directly representing the direction in a linear or logarithmic manner. The linear or logarithmic scale is selected according to application requirements. The directional target interval refers to the interval corresponding to the second quantization data. For example, the directional target interval corresponding to the second quantization data s is s1, and s1 is represented based on the second encoded data. In this case, the directional quantization data can also be represented using only less data, thereby reducing the amount of data transmitted.

[0038] In one embodiment, the feature data transmission method further includes: Step S201: Obtain the distribution range of the quantization data, wherein the quantization data includes first quantization data regarding amplitude and second quantization data regarding direction; Step S202: A preset quantization range interval is obtained by uniformly dividing the distribution range of the quantized data, or by dividing the distribution range of the quantized data by the amount of distributed data, wherein the amount of distributed data includes the distribution amount of the quantized data in different range intervals.

[0039] In this embodiment, the preset quantization range interval can use uniform quantization. For example, by obtaining the distribution range of quantized data, where the quantized data includes first quantized data regarding amplitude and second quantized data regarding direction, the range distribution of direction and amplitude quantized data can be determined based on a large number of experiments. Then, uniform scaling can be performed within the range distribution interval. For example, if the range distribution of both direction and amplitude quantized data is A1-A2, then A1-A2 can be divided into M intervals of equal size. Alternatively, non-uniform quantization levels can be used based on the feature distribution. For example, for amplitude, most features may be concentrated within a relatively small range, so a denser quantization level can be set within that range, while a coarser quantization scale can be used for rare large amplitude values, thereby reducing the overall quantization error without increasing the bit budget. A similar strategy can be used for direction quantization to improve the quantization accuracy of frequently occurring directions.

[0040] For example, assuming the quantized data distribution range of A1-A2 has a larger amount of data in A1-A3, then A1-A3 can be divided into more smaller intervals, while the distribution data in A3-A2 has a smaller amount of data, then A3-A2 can be divided into fewer larger intervals, or even into a single interval. Once the quantization range interval is determined, this interval can be directly called upon, and the interval can be directly represented using bit encoding. This flexibility allows the system to achieve the best compression efficiency and performance balance through strategy adjustments even with limited total bandwidth. Compared to a fixed uniform bit allocation scheme, this application can make fuller use of the available bit rate, compressing the data to a lower level while ensuring task performance. For example, the bit encoding representation can be either directly using bit encoding to represent the interval or directly representing the order of the intervals. In the case of the order representation, a feature is needed to indicate the position of the first, last, or one of the intervals. Of course, other representation methods can also be used, which will not be described in detail here.

[0041] Furthermore, based on the first, second, and / or third embodiments of this application applied to a vehicle control terminal described above, a fourth embodiment of the feature data transmission method of this application is proposed. In this embodiment, after the step of obtaining the initial feature data to be transmitted, the feature data transmission method further includes: Step S101: Obtain the current transmission environment at the current moment, wherein the current transmission environment includes the real-time network status and transmission task requirements; Step S102: If the real-time network status and transmission task requirements meet the preset transmission conditions, then the initial feature data is transformed by the preset first dimensionality reduction parameter, and the dimensionality-reduced feature vector after the dimensionality reduction transformation is used as the initial feature data. Furthermore, the feature data transmission method also includes: Step S103: If the real-time network status or transmission task requirements do not meet the preset transmission conditions, the initial feature data is transformed by the preset second dimensionality reduction parameter, and the dimensionality-reduced feature vector after the dimensionality reduction transformation is used as the initial feature data. The target dimension of the second dimensionality reduction parameter is smaller than the target dimension of the first dimensionality reduction parameter.

[0042] In this embodiment, after receiving the initial feature data to be transmitted, the initial feature data can be dimensionality-reduced to reduce the amount of data transmitted, for example, by directly reducing the initial feature data to the required dimension. For example, the entire dimensionality reduction process can also involve obtaining the current transmission environment at the current moment, where the current transmission environment includes the real-time network status and transmission task requirements. The real-time network status refers to the current network condition for communication with the cloud, and the transmission task requirements refer to the transmission requirements, such as high-precision transmission. Then, if it is determined that both the real-time network status and the transmission task requirements meet the preset transmission conditions, the initial feature data is dimensionality-reduced using a preset first dimensionality reduction parameter, and the dimensionality-reduced feature vector is used as the initial feature data. For example, when bandwidth is limited and the transmission task requirements are not high, the quantization precision can be reduced accordingly to reduce the amount of data transmitted. This adaptive adjustment can achieve the optimal trade-off between compression efficiency and task performance, i.e., using a lower dimension for transformation. The preset transmission conditions can be bandwidth limitation and low transmission precision requirements. These two conditions can be judged using their respective thresholds, such as determining bandwidth limitation when the rate is less than D, and determining low precision requirements when the precision is less than G, i.e., a lower dimension can be used for processing in this case. In another embodiment, when the real-time network status or transmission task requirements do not meet the preset transmission conditions, the initial feature data is transformed using a preset second dimensionality reduction parameter, and the dimensionality-reduced feature vector is used as the initial feature data. The target dimension of the second dimensionality reduction parameter is smaller than the target dimension of the first dimensionality reduction parameter. When bandwidth is ample or the transmission task requirements are high, the number of bits for amplitude quantization or the number of directional quantization levels can be increased to obtain a more refined reconstruction. For example, 10-dimensional data is normally used for subsequent processing; 6-dimensional data can be used in some cases; and 15-dimensional data can be used in high-requirement situations, thus greatly adapting to the needs and scenarios of data transmission.

[0043] For example, during the training phase of the dimensionality reduction model, the quantization encoding process of this application can be integrated into end-to-end training. Gradient propagation can be maintained through techniques such as pass-through estimation, enabling the feature extraction network to learn and generate feature representations that are more compatible with the method of this application. This approach is expected to further reduce the accuracy loss introduced by quantization, making the performance of the edge-cloud collaborative system closer to the original model.

[0044] Furthermore, based on the first, second, third, and / or fourth embodiments of this application applied to a vehicle control terminal described above, a fifth embodiment of the feature data transmission method of this application is proposed. In this embodiment, after the step of determining the target transmission data based on amplitude quantization data and direction quantization data, the method includes: Step a: Determine the amplitude scalar data corresponding to the amplitude quantization data, and determine the unit direction data corresponding to the direction quantization data; Step b: Determine the product between the amplitude scalar data and the unit direction data as the reconstructed feature data for the transmission.

[0045] In this embodiment, when the amplitude quantization data and directional quantification data The data are combined to form the final compressed feature representation, which is then packaged into a bitstream and sent to the cloud over the network. For example, the bits representing the amplitude quantization level index are sent first, followed by the bits representing each quantization level in sequence. The quantization index bits, such as Figure 3 The data is transmitted through feature-encapsulated code streams, and then decoded and reconstructed in the cloud. Reconstruction then takes place in the cloud. For example, the directional quantization and amplitude quantization functions are used to reverse-process the directional quantization and amplitude quantization data to obtain amplitude scalar data and unit direction data, which are then directly parsed to obtain the amplitude scalar data. and unit direction data Reconstruct feature data This is an approximate reconstruction of the original feature El. It also contains the approximate amplitude and orientation information of the original features, which can be directly used for subsequent visual multimodal task inference. That is, the reconstructed feature data retains the overall amplitude and orientation pattern of the features. Therefore, the reconstructed features can accurately represent the original features, making the performance of the cloud model close to that of horizontal cloud decoding and reconstruction using uncompressed features. Thus, the accuracy of feature data transmission is guaranteed through the above methods.

[0046] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the feature data transmission method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0047] This application also provides a feature data transmission device, referring to... Figure 4 Feature data transmission device: The data acquisition module A10 is used to acquire the initial feature data to be transmitted and determine the amplitude scalar data based on the initial feature data. The data quantization module A20 is used to perform amplitude quantization based on amplitude scalar data to obtain amplitude quantized data, and to perform direction quantization based on initial feature data and amplitude scalar data to obtain direction quantized data; The data transmission module A30 is used to determine the target transmission data based on the amplitude quantization data and the direction quantization data in order to realize the transmission of characteristic data.

[0048] The feature data transmission system provided in this application, employing the feature data transmission method described in the above embodiments, can solve the technical problem of low accuracy in feature data transmission. Compared with the prior art, the beneficial effects of the feature data transmission system provided in this application are the same as those of the feature data transmission method described in the above embodiments, and other technical features of the feature data transmission system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0049] This application provides a feature data transmission device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the feature data transmission method in Embodiment 1 above.

[0050] The following is for reference. Figure 5 The diagram illustrates a structural schematic of a feature data transmission device suitable for implementing the embodiments of this application. The feature data transmission device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The illustrated feature data transmission device is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0051] like Figure 5As shown, the feature data transmission device may include a processing system 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage system 1003 into random access memory (RAM) 1004. RAM 1004 also stores various programs and data required for the operation of the feature data transmission device. The processing system 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to input / output interface 1006: input system 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output system 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage system 1003 including, for example, magnetic tape, hard disk, etc.; and communication system 1009. Communication system 1009 allows the featured data transmission device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows featured data transmission devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0052] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication system, or installed from storage system 1003, or installed from read-only memory 1002. When the computer program is executed by processing system 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0053] The feature data transmission device provided in this application, employing the feature data transmission method described in the above embodiments, can solve the technical problem of low accuracy in feature data transmission. Compared with the prior art, the beneficial effects of the feature data transmission device provided in this application are the same as those of the feature data transmission method described in the above embodiments, and other technical features in this feature data transmission device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0054] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0055] The above are merely specific embodiments 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.

[0056] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the feature data transmission method described in the above embodiments.

[0057] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0058] The aforementioned computer-readable storage medium may be included in the feature data transmission device; or it may exist independently and not assembled into the feature data transmission device.

[0059] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the feature data transmission device, cause the feature data transmission device to: Acquire the initial feature data to be transmitted, and determine the amplitude scalar data based on the initial feature data; Amplitude quantization is performed on the amplitude scalar data to obtain amplitude quantized data, and direction quantization is performed on the initial feature data and amplitude scalar data to obtain direction quantized data; The target transmission data is determined based on the amplitude quantization data and the direction quantization data in order to achieve characteristic data transmission.

[0060] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0061] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0062] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0063] The computer-readable storage medium provided in this application stores computer-readable program instructions (i.e., a computer program) for executing the above-described feature data transmission method, thereby solving the technical problem of low accuracy in feature data transmission. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the feature data transmission method provided in the above embodiments, and will not be repeated here.

[0064] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the feature data transmission method described above.

[0065] The computer program product provided in this application can solve the technical problem of low accuracy in feature data transmission. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the feature data transmission method provided in the above embodiments, and will not be repeated here.

[0066] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for transmitting feature data, characterized in that, The feature data transmission method includes: Acquire the initial feature data to be transmitted, and determine the amplitude scalar data based on the initial feature data; Amplitude quantization is performed on the amplitude scalar data to obtain amplitude quantized data, and direction quantization is performed on the initial feature data and the amplitude scalar data to obtain direction quantized data; The target transmission data is determined based on the amplitude quantization data and the direction quantization data to achieve characteristic data transmission.

2. The feature data transmission method as described in claim 1, characterized in that, The step of obtaining amplitude quantized data by performing amplitude quantization based on the amplitude scalar data includes: The amplitude scalar data is quantized based on a preset amplitude quantization function to obtain multiple first quantized data, and the amplitude target range of each first quantized data is determined in a preset quantization range. Determine the first encoded data corresponding to each amplitude target interval, and use the first encoded data as amplitude quantization data.

3. The feature data transmission method as described in claim 2, characterized in that, The step of determining the first encoded data corresponding to each amplitude target interval includes: The number of intervals in the amplitude target interval is determined, and each amplitude target interval is encoded based on the target code corresponding to the number of intervals to obtain the first encoded data.

4. The feature data transmission method as described in claim 1, characterized in that, The step of obtaining direction-quantized data by performing direction quantization based on the initial feature data and the amplitude scalar data includes: The ratio between the initial feature data and the amplitude scalar data is determined as the unit direction data, and the unit direction data is quantized based on a preset direction quantization function to obtain multiple second quantized data. The direction target interval of each second quantized data in the preset quantization range interval is determined. Determine the second encoded data corresponding to each of the stated directional target intervals, and use the second encoded data as directional quantization data.

5. The feature data transmission method as described in any one of claims 2 or 4, characterized in that, The feature data transmission method further includes: Obtain the distribution range of the quantized data, wherein the quantized data includes first quantized data regarding amplitude and second quantized data regarding direction; A preset quantization range interval is obtained by uniformly dividing the distribution range of the quantized data, or by dividing the distribution range of the quantized data by the amount of distributed data, wherein the amount of distributed data includes the distribution amount of the quantized data in different range intervals.

6. The feature data transmission method as described in claim 1, characterized in that, After the step of acquiring the initial feature data to be transmitted, the feature data transmission method includes: Obtain the current transmission environment at the current moment, wherein the current transmission environment includes real-time network status and transmission task requirements; If both the real-time network status and the transmission task requirements meet the preset transmission conditions, then the initial feature data is subjected to a dimensionality reduction transformation using a preset first dimensionality reduction parameter, and the dimensionality-reduced feature vector after the dimensionality reduction transformation is used as the initial feature data.

7. The feature data transmission method as described in claim 6, characterized in that, The feature data transmission method further includes: If the real-time network status or the transmission task requirements do not meet the preset transmission conditions, the initial feature data is transformed by a preset second dimensionality reduction parameter, and the dimensionality-reduced feature vector after the dimensionality reduction transformation is used as the initial feature data, wherein the target dimension of the second dimensionality reduction parameter is smaller than the target dimension of the first dimensionality reduction parameter.

8. A feature data transmission device, characterized in that, The feature data transmission device includes: The data acquisition module is used to acquire the initial feature data to be transmitted and determine the amplitude scalar data based on the initial feature data. The data quantization module is used to perform amplitude quantization based on the amplitude scalar data to obtain amplitude quantized data, and to perform direction quantization based on the initial feature data and the amplitude scalar data to obtain direction quantized data; The data transmission module is used to determine the target transmission data based on the amplitude quantization data and the direction quantization data, so as to realize the characteristic data transmission.

9. A feature data transmission device, characterized in that, The feature data transmission device includes a processor and a memory. The memory stores a feature data transmission method program that can run on the processor. When the feature data transmission method program is executed by the processor, it implements the steps of the feature data transmission method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a feature data transmission method program, wherein when the feature data transmission method program is executed by a processor, it implements the steps of the feature data transmission method as described in any one of claims 1 to 7.

Images