Data compression method and device, electronic equipment and computer readable storage medium
Through predicted neighborhood models and optimization parameters based on data mechanisms, the problem of low data compression rate of high-precision mechanisms in the existing technology is solved, and efficient and reversible data compression and decompression is achieved, which is suitable for scientific calculation and remote sensing data.
Patent Information
- Application Number
- CN202310644160.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2024-12-03
AI Technical Summary
When existing data compression algorithms process high-precision mechanism data, the compression rate is low and the performance is poor. Especially for scientific computing data and high-precision remote sensing data that meet clear physical laws, general compression methods are difficult to effectively compress, resulting in high data storage costs.
Using a predictive neighborhood model based on data mechanism, by generating a prediction model for multi-dimensional data, optimizing model parameters, determining model errors, and storing model errors, parameters and boundary elements, the local correlation and linear characteristics of the data are used to achieve efficient Data compression and provide a reversible reconstruction scheme.
It realizes high compression rate and low model error, avoids the low reliability and large computing problems caused by complex models, and is suitable for data compression in complex scenarios, especially in scientific computing and high-precision remote sensing data, which significantly improves the compression effect. .
Smart Images

Figure CN119070828A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present disclosure relate to the field of computer technology, and more specifically, to the field of data compression technology. The embodiments of the present disclosure relate to a method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product for data compression and decompression. Background Art
[0002] In data storage, a large amount of data has specific physical mechanisms due to its generation and collection mechanism. High-performance computing (HPC) scientific computing data, such as turbulence, atmosphere and ocean, and other physical field numerical simulation data, correspond to clear model mechanisms (control equations); various observation data, especially geophysical satellite remote sensing data, such as sea surface height, temperature, atmospheric temperature and pressure, etc., whose evolution satisfies various conservation laws, also contain mechanisms; other industrial and social production data, such as those that conform to internal laws such as continuity and conservation, also have their specific mechanisms. Mining the data mechanism can greatly improve the compression rate of such data, so an effective data compression method based on data mechanism is needed. Summary of the invention
[0003] The embodiments of the present disclosure provide a data compression solution based on a data mechanism.
[0004] According to the first aspect of the present disclosure, a method for data compression is provided. The method includes: generating a prediction neighborhood model for the multidimensional data to be compressed; optimizing the model parameters of the prediction neighborhood model based on the multidimensional data; determining the model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, wherein the model error is associated with the internal elements of the multidimensional data; and storing the model error, the optimized model parameters, and the boundary elements of the multidimensional data. Based on this approach, the embodiments of the present disclosure fully extract the data mechanism for the characteristics of the mechanism data, and realize a high compression rate and high-performance compression algorithm with few model parameters, low model deviation, and few data support sets, and realizes reversible and stable reconstruction, avoiding the problems of low reliability and large amount of calculation caused by complex models.
[0005] In some embodiments of the first aspect, the model parameters of the prediction neighborhood model may indicate that an element of the multidimensional data is associated with at least one neighboring element of the element. Based on this approach, the local correlation of the data is utilized to simplify the model and reduce the computational complexity.
[0006] In some embodiments of the first aspect, the multidimensional data may be two-dimensional data, and the prediction neighborhood model may include a template covering elements in the two-dimensional data. Based on this approach, the data mechanism of a second-order or higher-order differential equation in a discrete form can be characterized, achieving a higher data compression rate.
[0007] In some embodiments of the first aspect, the method may further include: obtaining multidimensional data by partitioning the input data into blocks, wherein the multidimensional data is a portion of the partitioned input data. Based on this approach, the local linear characteristics of the data are utilized to reduce the model error and achieve a higher data compression rate.
[0008] In some embodiments of the first aspect, the method may further include: generating multidimensional data and a precision truncation error that conforms to the data precision from the input data according to the data precision; and storing the precision truncation error. Based on this approach, flexible data compression can be achieved by considering hardware performance and compression parameters.
[0009] In some embodiments of the first aspect, generating a prediction neighborhood model for multidimensional data to be compressed may include: determining multiple candidate prediction neighborhood models with different model parameters; determining a model error of the multidimensional data relative to each candidate prediction neighborhood model; and determining the candidate prediction neighborhood model with the smallest model error as the prediction neighborhood model for the multidimensional data. Based on this approach, an initial guess can be made for the prediction model, thereby improving the speed and performance of subsequent model optimization.
[0010] In some embodiments of the first aspect, the plurality of candidate prediction neighborhood models may include at least one of the following: an elliptic type difference operator; a hyperbolic type difference operator; or a parabolic type difference operator. Based on this approach, the differential equation type that is most likely to conform to the data mechanism is selected.
[0011] In some embodiments of the first aspect, the method may further include: generating preprocessed multidimensional data from the multidimensional data based on a preprocessing template for noise reduction, and optimizing the model parameters of the prediction neighborhood model may include: optimizing the model parameters of the prediction neighborhood model based on the preprocessed multidimensional data. Based on this approach, noise that is difficult to compress in the original data can be filtered out, thereby improving data compression efficiency.
[0012] In some embodiments of the first aspect, optimizing the model parameters of the prediction neighborhood model includes performing the following actions at least once: determining the predicted values of at least some elements of the multidimensional data based on the model parameters of the prediction neighborhood model; and updating the model parameters by minimizing the difference between the actual values and the predicted values of at least some elements. Based on this approach, the mechanism can be learned from the data, the model error can be reduced, and the compression rate can be improved.
[0013] In some embodiments of the first aspect, determining a model error of a prediction neighborhood model may include: determining a predicted value of an internal element of the multidimensional data based on an optimized model parameter and a boundary element of the multidimensional data; and determining a model error based on a difference between an actual value of the internal element and a predicted value. In some embodiments of the first aspect, determining a model error may include: obtaining a truncated model error for storage based on data accuracy. Based on this approach, flexible compression requirements may be met.
[0014] In some embodiments of the first aspect, storing boundary elements of multidimensional data may include: determining boundary elements of the multidimensional data to be stored based on optimized model parameters; and storing the determined boundary elements. In some embodiments of the first aspect, determining boundary elements of the multidimensional data to be stored may include: determining a layer of boundary elements of the multidimensional data for storage in response to the optimized model parameters corresponding to a non-singular matrix template; or determining a second layer of boundary elements of the multidimensional data for storage in response to the optimized model parameters corresponding to a singular matrix template. Based on this approach, a reversible and stable reconstruction can be obtained according to the local and equation well-posedness theory.
[0015] In some embodiments of the first aspect, the method may further include: evaluating the quality of the prediction neighborhood model and the multidimensional data based on the distribution of model errors. Based on this approach, the quality of the observation data, scientific computing data, and model quality can be evaluated from the perspective of data mechanism.
[0016] In some embodiments of the first aspect, the multidimensional data may include mechanistic data.
[0017] According to a second aspect of the present disclosure, a method for data decompression is provided. The method comprises: obtaining boundary elements of stored multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data; determining a set of equations for the multidimensional data based on the model parameters, elements, and model errors; and determining the internal elements of the multidimensional data by solving the set of equations.
[0018] In some embodiments of the second aspect, a model parameter of the predictive domain model may indicate that an element of the multidimensional data is associated with at least one neighboring element of the element.
[0019] In some embodiments of the second aspect, the multidimensional data may be two-dimensional data, and the predictive neighborhood model may include a template covering elements in the two-dimensional data.
[0020] In some embodiments of the second aspect, the method may further include: obtaining a precision truncation error of the stored multidimensional data; and accumulating internal elements and boundary elements of the multidimensional data with the corresponding precision truncation error.
[0021] In some embodiments of the second aspect, determining a system of equations for internal elements of the multidimensional data may include: determining a matrix of the system of equations based on model parameters; and determining components of a column vector of the system of equations based on boundary elements and model errors.
[0022] In some embodiments of the second aspect, solving the system of equations may include: determining a manner of solving the system of equations based on a type of the matrix; and solving the system of equations using the determined manner.
[0023] In some embodiments of the second aspect, when the type of the matrix is a non-singular matrix, the boundary elements of the stored multidimensional data include one layer of boundary elements of the multidimensional data; or when the type of the matrix is a singular matrix, the boundary elements of the stored multidimensional data include two layers of boundary elements of the multidimensional data.
[0024] In some embodiments of the second aspect, the multidimensional data includes mechanistic data.
[0025] According to a third aspect of the present disclosure, an electronic device is provided. The electronic device includes: a processing unit and a memory, the processing unit executes instructions in the memory, so that the electronic device executes a method, the method including: generating a prediction neighborhood model for multidimensional data to be compressed; optimizing model parameters of the prediction neighborhood model based on the multidimensional data; determining a model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, the model error being associated with internal elements of the multidimensional data; and storing the model error, the optimized model parameters, and the boundary elements of the multidimensional data.
[0026] According to a fourth aspect of the present disclosure, an electronic device is provided. The electronic device includes: a processing unit and a memory, the processing unit executes instructions in the memory, so that the electronic device performs a method, the method including: obtaining boundary elements of stored multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data; determining a set of equations for the multidimensional data based on the model parameters, elements, and model errors; and determining the internal elements of the multidimensional data by solving the set of equations.
[0027] According to a fifth aspect of the present disclosure, a device for data compression is provided. The device includes: a model generation unit configured to generate a prediction neighborhood model for multidimensional data to be compressed; a model optimization unit configured to optimize the model parameters of the prediction neighborhood model based on the multidimensional data; a model error determination unit configured to determine the model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, the model error being associated with the internal elements of the multidimensional data; and a data storage unit configured to store the model error, the optimized model parameters, and the boundary elements of the multidimensional data.
[0028] According to a sixth aspect of the present disclosure, a device for data decompression is provided. The device includes: an acquisition unit configured to acquire boundary elements of stored multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data; an equation group determination unit configured to determine an equation group for the multidimensional data based on the model parameters, the boundary elements, and the model errors; and a solution unit configured to determine the internal elements of the multidimensional data by solving the equation group.
[0029] According to a seventh aspect of the present disclosure, a computer-readable storage medium is provided, on which one or more computer instructions are stored, wherein the one or more computer instructions are executed by a processor to enable the processor to execute the method according to the first aspect or the second aspect of the present disclosure.
[0030] According to an eighth aspect of the present disclosure, a computer program product is provided, comprising machine executable instructions, which, when executed by a device, enable the device to perform the method according to the first aspect or the second aspect of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:
[0032] Figure 1 A schematic diagram showing the generation of mechanistic data;
[0033] Figure 2 A schematic diagram showing an example computing device capable of implementing embodiments of the present disclosure;
[0034] Figure 3 A schematic flow chart of a process for data compression according to an embodiment of the present disclosure is shown;
[0035] Figure 4 A schematic flow chart of a process for data decompression according to an embodiment of the present disclosure is shown;
[0036] Figure 5A and 5B An exemplary storage format of mechanism data is shown;
[0037] Figure 6 A schematic flow chart showing another process for data compression according to an embodiment of the present disclosure is shown;
[0038] Figure 7 A schematic flow chart showing another process for data decompression according to an embodiment of the present disclosure is shown;
[0039] Fig. 8A A schematic diagram of data compression for accurate solution of elliptic equations according to an embodiment of the present disclosure is shown;
[0040] Figure 8B A schematic diagram showing data compression of a numerical solution of the NS equation according to an embodiment of the present disclosure is shown;
[0041] Fig.9A shows an exemplary distribution of sea surface temperature remote sensing data before and after compression according to an embodiment of the present disclosure;
[0042] Fig. 9B An exemplary mechanism extraction effect of sea surface temperature remote sensing data according to an embodiment of the present disclosure is shown;
[0043] Fig.10 A schematic block diagram showing an apparatus for data compression according to an embodiment of the present disclosure; and
[0044] Fig.11 A schematic block diagram of a device for data decompression according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION
[0045] The technical solution of the present disclosure will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments.
[0046] The technical solutions in the embodiments of the present disclosure will be described below in conjunction with the accompanying drawings in the embodiments of the present disclosure. In the description of the embodiments of the present disclosure, unless otherwise specified, " / " means and / or. For example, A / B can represent A or B, or A and B. "And / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present disclosure, "plurality" or "multiple" refers to two or more than two.
[0047] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this embodiment, unless otherwise specified, "plurality" means two or more.
[0048] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to be limitations on the present disclosure. As used in the specification and appended claims of the present disclosure, the singular expressions "one", "a kind", "said", "above", "the" and "this" are intended to also include expressions such as "one or more", unless there is a clear contrary indication in the context. It should also be understood that in the following embodiments of the present disclosure, "at least one", "one or more" refer to one, two or more. The term "and / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist; for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.
[0049] References to "one embodiment" or "some embodiments" etc. described in this specification mean that a particular feature, structure or characteristic described in conjunction with the embodiment is included in one or more embodiments of the present disclosure. Thus, the phrases "one embodiment", "some embodiments", "another embodiment", "some other embodiments", etc. appearing in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.
[0050] The numbers or numerical values used in this specification are all illustrative and are only intended to facilitate understanding of the techniques of the embodiments of the present disclosure, and are by no means intended to limit the scope of the present disclosure.
[0051] Overview
[0052] In data storage, a large amount of data has specific physical mechanisms due to its generation and collection mechanism. Scientific computing data of high performance computing (HPC), such as numerical simulation data of various physical fields such as turbulence, atmosphere and ocean, correspond to clear model mechanisms (for example, control equations); various types of observation data, especially geophysical satellite remote sensing data, such as sea surface height, temperature, atmospheric temperature and pressure, etc., whose evolution satisfies various conservation laws, also contain mechanisms; other industrial and social production data, such as those that conform to internal laws such as continuity and conservation, also have their specific mechanisms. Mining data mechanisms can help improve the compression rate of data. The inventors note that mechanism data can usually be characterized by differential equations, and the solutions of differential equations can be determined by appropriate initial values or boundary conditions and source terms. Therefore, compared with general compression methods based on empirical models, data compression methods based on high-order mechanism models have unique advantages for the above scenarios.
[0053] refer to Figure 1 , which shows a schematic diagram of the generation of mechanism data. In general, according to the storage item 150 of the mechanism data, the model data 120 is obtained from the physical law control equation 110, and then the observation data 1230 is generated from the model data 120. The storage item 150 of the mechanism data may include two parts: a data support 160 and a data error 170. The data support 160 may include a well-posedness condition 162 and a model parameter 164. The well-posedness condition 162 may be the initial value and boundary value of the observed physical field. The model parameter 164 reflects the connection or dependency between the data affected by the physical law control equation 110. The data error 170 may include a model error 172 and a calculation error 174. The model error 172 may be caused by model approximation or noise. When a differential equation is used to characterize the mechanism data, the model error may be expressed as a source term of the differential equation (herein, the model error and the source term may be used interchangeably). The calculation error 174 may be an error caused by data accuracy, for example, discretization or rounding. When generating mechanism data, firstly, the control equation 110 is controlled according to the physical law, and the model data 120 is generated in combination with the data support 160. Furthermore, the model data 120 can be superimposed with the data error 170 to obtain the observation data 130. The compression effect and generalization can be further optimized by reconstructing a mechanism that is more consistent with the data based on the idea of learning. From the perspective of data storage, when the model can better reflect the physical laws, the storage overhead of the model error 172 caused by the model approximation deviation can be reduced. This requires selecting the appropriate approximation order, control equation type, model parameters, etc. as much as possible.
[0054] There are some schemes based on fixed predictors and entropy coding to achieve data compression. They perform basis transformation on the data in the spatial domain, remove spatial redundancy, and use the sparsity of frequency domain coefficients to divide them into different bit planes for encoding to save storage space. Common predictors include Lorenzo predictors and interpolation predictors. However, this fixed empirical model has limited applicability to the compression effect of scientific computing data in complex scenes. It has a good compression effect on data that meets one type of mechanism, but for data of another type of mechanism, the model may not be able to well describe the physical laws, resulting in large deviations, resulting in large model errors and limited compression effects. However, the data characteristics of complex scenes in scientific computing data are often quite different, and specific models need to be designed according to the data characteristics to improve the compression rate, which brings new challenges.
[0055] There are also some solutions that use neural networks to approximate data, determine and store network parameters and data residuals through optimization. Due to the complexity and nonlinearity of various neural network models, the interpretability is weak, and it is not convenient to analyze the compression results. This type of method focuses on general data approximation and does not involve physical law control equations. Therefore, for data scenarios with mechanisms, it often cannot well reflect the differences in data features and generate intrinsic data support sets. When the degree of approximation is high, the storage cost of model parameters is high, and when the degree of approximation is low, the storage volume of model errors is large, so the comprehensive compression effect is limited.
[0056] In view of the fact that the existing general data compression algorithms and image compression algorithms have problems such as low data pertinence, low compression rate, and low performance on data with clear mechanisms such as scientific computing data and high-precision remote sensing data, the embodiments of the present disclosure provide a data compression scheme that can fully extract the data mechanism based on the characteristics of the mechanism data. By utilizing the unique characteristics and mechanism characteristics of floating-point data with mechanisms such as scientific data and high-precision remote sensing data compared to general data and image data, the proposed data compression scheme has fewer model parameters and lower model errors compared to existing schemes, and thus has a higher data compression rate. The embodiments of the present disclosure also provide a corresponding data decompression scheme, which can adaptively obtain a mechanism model with data pertinence based on the data, and the obtained model can achieve reversible and stable reconstruction when used for data compression, avoiding problems such as low reliability and large amount of calculation caused by complex models. In short, the embodiments of the present disclosure adopt a learning method to mine the implicit mechanism behind the data in a data-driven manner, optimize the model parameters, and are suitable for complex scenarios.
[0057] Example Environment
[0058] Figure 2 Schematic diagram of an example computing device 200 capable of implementing embodiments of the present disclosure is shown. It should be understood that Figure 2 The computing device 200 shown is merely exemplary and should not constitute any limitation on the functionality and scope of the implementation described in the present disclosure. Figure 2 As shown, components of computing device 200 may include, but are not limited to, one or more processors or processing units 210 , memory 220 , storage device 230 , one or more communication units 240 , one or more input devices 250 , and one or more output devices 260 .
[0059] In some implementations, the computing device 200 may be implemented as various user terminals or service terminals with computing capabilities. The service terminal may be a server, a large computing device, etc. provided by various service providers. The user terminal may be any type of mobile terminal, fixed terminal, or portable terminal, including a mobile phone, a site, a unit, a device, a multimedia computer, a multimedia tablet, an Internet node, a communicator, a desktop computer, a laptop computer, a notebook computer, a netbook computer, a tablet computer, a personal communication system (PCS) device, a personal navigation device, a personal digital assistant (PDA), an audio / video player, a digital camera / camcorder, a positioning device, a television receiver, a radio broadcast receiver, an electronic book device, a game device, or any combination thereof, including accessories and peripherals of these devices or any combination thereof.
[0060] Processing unit 210 may be a real or virtual processor and may be capable of performing various processes according to a program stored in memory 220. In a multi-processor system, multiple processing units execute computer executable instructions in parallel to increase the parallel processing capabilities of computing device 200. Processing unit 210 may also be referred to as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor, a controller, or a microcontroller.
[0061] The computing device 200 typically includes a plurality of computer storage media. Such media may be any available media accessible to the computing device 200, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 220 may be a volatile memory (e.g., a register, a cache, a random access memory (RAM)), a non-volatile memory (e.g., a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. The memory 220 may include a compression / decompression tool 222 implemented as a program module, and the compression / decompression tool 222 may be configured to perform data compression and data decompression as described herein. The compression / decompression tool 222 may be accessed and run by the processing unit 110 to implement corresponding data compression and data decompression functions.
[0062] Storage device 230 may be a removable or non-removable medium and may include machine-readable media that can be used to store information and / or data and can be accessed within computing device 200. Computing device 200 may further include additional removable / non-removable, volatile / non-volatile storage media. Although not described herein, Figure 1 As shown in , a disk drive for reading or writing from a removable, nonvolatile disk and an optical drive for reading or writing from a removable, nonvolatile optical disk can be provided. In these cases, each drive can be connected to a bus (not shown) by one or more data media interfaces.
[0063] The communication unit 240 enables communication with another computing device via a communication medium. Additionally, the functions of the components of the computing device 200 can be implemented in a single computing cluster or multiple computing machines that can communicate via a communication connection. Therefore, the computing device 200 can operate in a networked environment using a logical connection with one or more other servers, a personal computer (PC), or another general network node.
[0064] Input device 250 may be one or more various input devices, such as a mouse, keyboard, trackball, voice input device, etc. Output device 260 may be one or more output devices, such as a display, a speaker, a printer, etc. Computing device 200 may also communicate with one or more external devices (not shown) through communication unit 240 as needed, such as storage devices, display devices, etc., communicate with one or more devices that allow a user to interact with computing device 200, or communicate with any device that allows computing device 200 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).
[0065] In some implementations, in addition to being integrated on a single device, some or all of the various components of the computing device 200 can also be set in the form of a cloud computing architecture. In a cloud computing architecture, these components can be remotely arranged and can work together to implement the functions described in the present disclosure. In some implementations, cloud computing provides computing, software, data access and storage services, which do not require end users to know the physical location or configuration of the system or hardware that provides these services. In various implementations, cloud computing uses appropriate protocols to provide services through a wide area network (such as the Internet). For example, a cloud computing provider provides applications through a wide area network, and they can be accessed through a web browser or any other computing component. The software or components of the cloud computing architecture and the corresponding data can be stored on a server at a remote location. The computing resources in a cloud computing environment can be merged at a remote data center location or they can be dispersed. Cloud computing infrastructure can provide services through a shared data center, even if they appear as a single access point for users. Therefore, the components and functions described herein can be provided from a service provider at a remote location using a cloud computing architecture. Alternatively, they can also be provided from a traditional server, or they can be installed on a client device directly or otherwise.
[0066] The computing device 200 may implement data compression according to various implementations of the present disclosure. Figure 1As shown, the computing device 200 can receive data to be compressed 270 through the input device 250. The data to be compressed 270 can be, for example, observation data of a mechanism or data obtained through simulation. Alternatively, the computing device 200 can also read the data to be compressed 270 from the storage device 230 or receive the data to be compressed 270 from other devices from the communication device 240. The computing device 200 can transmit the data to be compressed 270 to the compression / decompression tool 222. The compression / decompression tool 222 generates corresponding compressed data 280 based on the data to be compressed 270. The compression can be lossless compression or lossy compression. The compressed data 280 can occupy less storage space than the original data to be compressed 270.
[0067] The embodiments of the present disclosure can be applied to unstructured data storage, for example, scientific computing floating-point format data storage. Exemplary unstructured data may include, but are not limited to: remote sensing images, cryo-electron microscopy, gene sequencing, energy exploration, ocean measurement, big data, etc. Among these unstructured data, scientific computing floating-point data accounts for a large proportion, and the embodiments of the present disclosure can achieve a large-scale reduction that is superior to traditional solutions, thereby reducing storage costs. In some implementations, the embodiments of the present disclosure can also be applied to high-precision satellite observation data storage. Various types of high-precision geophysical satellite remote sensing data, such as sea surface height, temperature, atmospheric temperature pressure, etc., their evolution is governed by various conservation laws, etc., have physical mechanisms and corresponding control equations, and are high-precision floating-point data. General compression methods such as traditional image data are no longer applicable, and the embodiments of the present disclosure can provide effective compression for such data.
[0068] According to an embodiment of the present disclosure, the data to be compressed 270 may be in the form of multidimensional data (e.g., dimensions such as space and time), and the computing device 200 uses a prediction neighborhood model to mine data mechanisms and achieve data compression. According to the prediction neighborhood model, elements in the multidimensional data may be considered to be associated with elements in its neighborhood. For example, the actual values of each element of the multidimensional data may satisfy a specific differential equation, so the elements therein may be obtained by a linear combination of its adjacent elements (e.g., the prediction neighborhood model is a nine-square grid template corresponding to a second-order neighborhood). Using the prediction model, the computing device 200 can calculate the predicted values of the internal elements from the boundary elements of the multidimensional data, and use the difference between the actual value and the predicted value of the internal element (model error) as the objective function for optimization to optimize the model parameters of the prediction neighborhood model. Thus, a model error as sparse as possible is obtained to improve the data compression rate. The computing device 200 can also encode (e.g., entropy encoding) the optimized model parameters, the edge elements of the multidimensional data, and the model error, thereby obtaining and storing the compressed data 280.
[0069] According to an embodiment of the present disclosure, the computing device 200 may also receive the compressed data obtained by the compression method according to an embodiment of the present disclosure and reconstruct the compressed data to obtain the decompressed data ( Figure 1 (not shown in the figure). The computing device 200 parses the compressed data to obtain the boundary elements of the multidimensional data to be decompressed, the model parameters of the prediction neighborhood model, and the model errors. Then, the computing device 200 can parse the data to construct a system of equations, and determine the internal elements of the multidimensional data by solving the combination of equations. Specifically, the matrix of the system of equations can be constructed from the model parameters of the prediction neighborhood model, and the column vectors of the system of equations can be constructed from the model errors and edge elements. Depending on whether the matrix is a singular matrix, there are different ways to solve the system of equations. If it is a non-singular matrix, the system of equations can be solved directly. If it is a singular matrix, it can be solved by solving the generalized inverse. Therefore, this data reconstruction is reversible and stable.
[0070] Data compression and decompression
[0071] The following references Figures 3 to 11 Embodiments of the present disclosure are further described.
[0072] Figure 3 1 is a schematic flow chart of a process 300 for data compression according to an embodiment of the present disclosure. The method 300 may be performed by, for example Figure 1 More specifically, the method 300 may be implemented by the computing device 200 shown in FIG. Figure 2 It should be understood that the method 300 may also include additional actions not shown and / or may omit the actions shown, and the scope of the present disclosure is not limited in this respect. Figure 2 The method 300 is described below.
[0073] like Figure 3As shown, in box 310, the computing device 200 generates a prediction neighborhood model for the multidimensional data to be compressed. For multidimensional data with mechanisms that can be characterized by differential equations, the prediction neighborhood model can indicate that elements in the multidimensional data are associated with their adjacent elements. The prediction neighborhood model can be used to predict an element from its adjacent elements. In some implementations, the prediction neighborhood model can cover the first-order, second-order or more order neighborhoods near a given element. The order of the neighborhood here can predict the size of the neighborhood model. For example, in the case where the multidimensional data is two-dimensional data, the prediction neighborhood model can be a template covering elements in the two-dimensional data. For example, if the second-order prediction model is used, the template is a 3x3 nine-square template (the central element corresponds to the predicted element). The prediction neighborhood model has model parameters, and the model parameters can be coefficients corresponding to elements in the template. In order to simplify model learning, facilitate theoretical analysis, and reduce computational complexity, the mechanism of multidimensional data is mined by local linearization. The model parameters of the prediction neighborhood model can be used as weights to calculate the predicted elements. In other words, an element in multidimensional data can be represented as a linear combination of its neighboring elements.
[0074] In some embodiments, considering the hardware performance and compression parameters of the computing device, the input data input to the computing device 200 can be divided into blocks, and each block is used as the multidimensional data to be compressed. As a result, the data can be better approximated locally by differential equations, thereby achieving a better compression effect. In some embodiments, the input data can also be truncated or rounded according to the data precision to generate multidimensional data that meets the data precision and the corresponding precision truncation error. The precision truncation error can be stored as the calculation error 174.
[0075] In some embodiments, the computing device 200 may also initialize the prediction neighborhood model. Initialization includes selecting appropriate model parameters to accelerate the subsequent optimization process and obtain a prediction model that more accurately reflects the data mechanism. In some implementations, the computing device 200 may select a candidate prediction neighborhood model from multiple candidate prediction neighborhood models with different preset model parameters. For example, when the multidimensional data to be compressed is a two-dimensional array and a second-order prediction model (corresponding to a mechanism that conforms to a second-order partial differential equation) is used, the candidate prediction neighborhood model may include an elliptic type difference operator, a hyperbolic type difference operator, and a parabolic type difference operator of size 3x3. The computing device 200 applies the candidate prediction neighborhood model to the multidimensional data to be compressed, and calculates the error size (e.g., source term modulus) between the actual value and the predicted value of the multidimensional data. Then, the computing device 200 may select the candidate prediction neighborhood model with the smallest model error and apply the corresponding model parameters as the initial guess. In order to simplify the calculation, the elements participating in the initial guess may be part of the multidimensional data but not necessarily all. Additionally or alternatively, the computing device 200 may also directly use pre-specified initial model parameters.
[0076] Since actual observation data often contains random noise, and random noise cannot be effectively compressed using a prediction model, it is beneficial to preprocess the multidimensional data to reduce noise. In some embodiments, the computing device 200 can use a preprocessing template to preprocess the multidimensional data. The preprocessing model can be, for example, a mean template. The noise reduction effect is achieved by calculating the average value of the elements covered by the template. Thus, processed multidimensional data is generated for subsequent model parameter optimization.
[0077] In box 320, the computing device 200 optimizes the model parameters of the prediction neighborhood model based on the multidimensional data. The optimization of model parameters can also be referred to as model learning. The computing device 200 learns the model parameters through an optimization algorithm based on the data. In some embodiments, the computing device 200 can use the local or overall model error size as the objective function of the optimization algorithm to optimize the model parameters. When using local optimization, the computing device 200 selects and traverses a part of the elements (e.g., data points outside the boundary) to add to the objective function. When using global optimization, the computing device 200 calculates the model errors of all elements. As mentioned above, the mechanism data can be characterized by differential equations, and the solution of the differential equation is determined by appropriate initial boundary values, so in the optimization process, the boundary elements of the multidimensional data are used to calculate the predicted values of the internal elements.
[0078] The optimization process can be performed in an iterative manner. In each iteration, the computing device 200 determines the predicted values of at least some elements of the multidimensional data based on the model parameters of the current prediction neighborhood model. In some implementations, the computing device 200 uses boundary elements to calculate the predicted values of these elements, and then updates the model parameters by minimizing the difference between the actual values and the predicted values of these elements (i.e., the model error). In the case where the prediction neighborhood model is a linear template, the calculation of the objective function only involves addition and multiplication operations, which has a small computational overhead.
[0079] In box 330, the computing device 200 determines the model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters. The model error is associated with the internal elements of the multidimensional data. After the model error optimization is completed, the model parameters of the prediction neighborhood model reflecting the data mechanism are determined. The prediction model and the quality of multidimensional data can be evaluated by analyzing the model error. It can be understood that when the model reflects the physical law well, the model error of each element calculated in box 330 is smaller than its predicted value (just as the objective function of optimization). At this time, a relatively sparse model error (mostly zero) can be obtained to achieve a higher compression rate. On the other hand, the distribution of the model error can also be judged. If the model error is close to a high-speed distribution with a mean of 0, it means that the data mechanism is well extracted and the quality of the compressed data is also good.
[0080] In some embodiments, the computing device 200 may obtain the truncated model error according to the data precision. The data precision may be a fixed preset number of bits, or may be further optimized during the compression process. In some embodiments, the computing device 200 considers directly encoding and storing the boundary elements of the multidimensional data, and therefore, it calculates the difference between each internal element and the predicted value as the model error of the predicted neighborhood model for storage.
[0081] Considering that the system of equations will be solved based on the prediction model during decompression, for reversible and stable decompression, the computing device 200 determines the boundary elements to be stored based on the model parameters of the optimized prediction neighborhood model. If the template of the optimized model parameters corresponds to a non-singular matrix template (i.e., full rank), the computing device 200 determines a layer of boundary elements (i.e., the outermost elements) of the multidimensional data for storage. If the optimized model parameters correspond to a singular matrix template, the computing device 200 determines a second layer of boundary elements (the outermost and second outermost elements) of the multidimensional data for storage. In this case, two layers of boundary elements will be used to solve the overdetermined system by a generalized inverse method.
[0082] In box 240, the computing device 200 stores the model error, the optimized model parameters, and the boundary elements of the multidimensional data. The computing device can store the above data items by, for example, entropy coding. Taking the prediction neighborhood model as a nine-grid template and the data to be compressed as two-dimensional data as an example, the computing device only needs to store 8 model parameters (the parameter of the central element can be fixed to 1 and does not need to be stored), the sparse model error (entropy coding helps to reduce storage space), and the boundary elements limited to one dimension (at most two layers), achieving a higher data compression rate.
[0083] Figure 4 4 shows a schematic flow chart of a process 400 for data decompression according to an embodiment of the present disclosure. The method 400 may be performed by, for example Figure 2 More specifically, the method 400 may be implemented by the computing device 200 shown in FIG. Figure 2 It should be understood that the method 400 may also include additional actions not shown and / or may omit the actions shown, and the scope of the present disclosure is not limited in this respect. Figure 2 To illustrate method 400 .
[0084] At block 410, the computing device 200 obtains the stored boundary elements of the multidimensional data, the model parameters of the prediction neighborhood model, and the model error of the prediction neighborhood model, wherein the model error is associated with the internal elements of the multidimensional data. The boundary elements, model parameters, and model errors are calculated during the data compression process (e.g., Figure 3 The process 300 shown in FIG. 2 is shown in FIG. 2 and is stored as compressed data. The computing device 200 receives the stored compressed data and extracts boundary elements, model parameters and model errors of the multidimensional data to be decompressed from the compressed data. The multidimensional data may be mechanism data that conforms to a physical law governing equation, such as a differential equation.
[0085] In some embodiments, the model parameter indicates that an element of the multidimensional data is associated with at least one adjacent element of the element. For example, using the prediction neighborhood model, the predicted value of a certain element can be calculated from the neighborhood element. The model error reflects the difference between the actual value of the internal element of the multidimensional data and the predicted value of the prediction neighborhood model. In some embodiments, the multidimensional data can be two-dimensional data, and the prediction neighborhood model includes a template covering the elements in the two-dimensional data.
[0086] At block 420, the computing device 200 determines a system of equations for the multidimensional data based on the model parameters, the elements, and the model errors. The system of equations may be a linear system of equations. In some embodiments, the computing device 200 may determine a matrix of the system of equations based on the model parameters, and a column vector of the system of equations based on the boundary elements and the model errors.
[0087] In box 430, the computing device 200 determines the internal elements of the multidimensional data by solving the system of equations. In some embodiments, the solution method depends on the type of the matrix of the system of equations. If the type of the matrix is a non-singular matrix, the computing device 200 can directly solve the linear equations to obtain the internal elements of the multidimensional data. In this case, the computing device 200 can determine that the boundary elements obtained in box 410 include a layer of boundary elements (i.e., the outermost elements) of the multidimensional data to be decompressed. If the type of the matrix is a singular matrix (i.e., ill-conditioned), the computing device 200 can solve the corresponding overdetermined system of equations by a generalized inverse method. In this case, the computing device 200 can determine that the boundary elements obtained in box 410 include two layers of boundary elements (e.g., the outermost and second outermost elements) of the multidimensional data to be decompressed.
[0088] In some embodiments, the compressed data may further include a precision truncation error. The computing device 200 may further obtain the stored precision truncation error and accumulate the determined internal elements and boundary elements of the multidimensional data with the corresponding precision truncation error. Thus, the computing device 200 obtains the decompressed multidimensional data.
[0089] The following reference figures 5 to Fig.11 Further embodiments of the present disclosure are described. Figure 5A and 5B An exemplary storage format of mechanism data is shown to help understand the compression method proposed in the present disclosure.
[0090] Assume that the two-dimensional data to be compressed satisfies the Laplace equation Δu=0, and the background is a steady-state temperature field, electric field, etc. If the data is a numerical solution to the above equation, for example, it can be calculated by the equally spaced second-order central difference method, then the data satisfies the equation:
[0091] u i,j-1 +u i,j+1 +u i-1,j +u i+1,j -4u i,j =0 (1)
[0092] That is, the data of the middle point can be determined by the values of its surrounding adjacent points:
[0093]
[0094] For data that meets this mechanism, compression can be achieved by storing data at intervals, such as Figure 5A As shown, blank data points can be predicted by surrounding data points without storing the original values.
[0095] Furthermore, the values of the data satisfying the above discrete equation at the internal points only depend on the values of the external boundaries. Let the matrix corresponding to the solver be A. Then, when storing, only the boundary data needs to be stored, and the internal data can be directly calculated, such as Figure 5B The internal data can be calculated from the boundary data by the following formula:
[0096] u interior =Au boundary (3)
[0097] For this discrete equation, the solver A is uniquely determined. In practical problems, if the data does not satisfy the above average property, that is, the above homogeneous discrete equation, the influence of the source term f needs to be considered, that is:
[0098] Δu=f (4)
[0099] In this case, the corresponding difference format is:
[0100]
[0101] The above difference equation can be written in matrix-vector form as
[0102] Ku interior =b (6)
[0103] Where K is the stiffness matrix, which is determined by the coefficients of the difference equation, and b depends on u boundary and f, that is, the internal data is determined by the boundary value u boundary and the source term f. When f is sparse and the data is sufficiently uniform, f is a low-value disturbance and requires less storage space.
[0104] In some embodiments, the data model is characterized by a differential equation. For a set of two-dimensional data, it is assumed that a certain differential equation is satisfied
[0105]
[0106] in is the difference operator, and its corresponding matrix form is
[0107]
[0108] in is the stiffness matrix corresponding to the difference operator, u bord is the "data support", i.e., the equation well-posedness data. In some embodiments, the differential operator is learned by data Or the corresponding stiffness matrix K, minimize the source term |f|, that is, desparse f to the greatest extent, reducing the information entropy / storage amount of f. In this case, only the "data support" and the sparse source terms need to be stored. Decompression is performed based on the matrix K during decompression. It should be noted that the embodiments of the present disclosure do not require the data to obey a certain differential format, and can also be applied to variable coefficient equations to achieve a certain compression effect.
[0109] In some implementations, specifically, let the difference operator By parameter θ = {C i} Determine, here we take the coefficient combination of 9 adjacent points as an example, i = 1, 2...8, the prediction neighborhood model is as follows:
[0110] <![CDATA[C1u i-1,j+1 ]]> <![CDATA[C2u i,j+1 ]]> <![CDATA[C3u i+1,j+1 ]]> <![CDATA[C4u i-1,j ]]> <![CDATA[u i,j ]]> <![CDATA[ C 5th i+1,j ]]> <![CDATA[C6u i-1,j-1 ]]> <![CDATA[C7u i,j-1 ]]> <![CDATA[C8u i+1,j-1 ]]>
[0111] In other words, we need to determine the parameters Make
[0112] u i,j =C1u i-1,j+1 +…+C8u i+1,j-1 +f i,j , (9)
[0113] and minimize f i,j Therefore, the learning of the operator (or the corresponding stiffness matrix) is transformed into solving the following optimization problem:
[0114]
[0115] in When this method is applied, in order to learn the differential operator (or stiffness matrix) corresponding to the mechanism, a batch of data can be selected In some special scenarios, in order to save algorithm overhead, only local data can be used to optimize the source item to obtain And the corresponding stiffness matrix Then encode and store the boundary value u boundary , source term and the model parameters θ.
[0116] During decompression, by solving
[0117]
[0118] Get u interior .
[0119] The disclosed solution is suitable for storing numerical solution data of differential equations. At the same time, it also has good effects on data with clear physical mechanisms, such as atmospheric and ocean observation data or industrial data. (difference operator), which actually corresponds to the mechanism described by the second-order differential equation, is general and computationally convenient and stable. It should be pointed out that the specific implementation of the above algorithm, including the form of "data support" and the stable reconstruction of the original data during decompression, is based on the theory of differential equations and is therefore not trivial.
[0120] Figure 6 FIG. 6 is a schematic flow chart of another process 600 for data compression according to an embodiment of the present disclosure. The process 600 may be a reference to Figure 3 An exemplary implementation of process 300 is described. It should be understood that one or more steps of process 600 may be omitted, and there may be more steps.
[0121] In response to receiving the input data to be compressed, at block 610, the computing device 200 performs data segmentation and initialization. The computing device 200 can segment the data according to the hardware processing capability and the compression parameters. The computing device 200 also sets the model parameters and the number of truncation bits for the subsequent optimization process according to the number of data precision bits (decimal or binary significant digits / absolute precision). The model error (i.e., source term) can be estimated by a preset candidate prediction model. Exemplary candidate prediction models are as follows:
[0122]
[0123] The candidate prediction models above correspond to elliptic, hyperbolic and parabolic difference operators respectively. The model with the smallest model error can be selected as the initial guess of the prediction model. Alternatively or additionally, the initial values of the model parameters can also be specified based on other prior tests.
[0124] In block 620, computing device 200 performs data preprocessing. In some embodiments, an average template may be used to reduce noise in the original data, because noisy data often cannot be effectively compressed using a mechanism model and may interfere with subsequent optimization processes.
[0125] In block 630, the computing device 200 optimizes the template parameters. The computing device 200 learns the data mechanism through the optimization algorithm. Taking the above-mentioned nine-square grid difference template as an example, the computing device 200 optimizes the template parameters (template coefficients) based on the multi-dimensional data after block division. An exemplary objective function can be:
[0126]
[0127] Where i, j traverse the data points outside the boundary, and the optimization process corresponds to minimizing the above objective function, which is an unconditional optimization. The calculation of the objective function only involves coefficient multiplication and algebraic summation process, so the computational overhead is small, and it is also convenient to generalize to the deployment of calculations on parallel computing devices (such as GPUs). In some implementations, the optimization algorithm can be the Levenberg-Marquardt algorithm, which takes into account both convergence speed and stability, and the initial value of the parameter is selected in box 610. Other optional parameters may include the number of iterations, the number of objective function calls, tolerance, etc. In some implementations, if global optimization is specified in the user settings, the SQP algorithm or other global optimization algorithms can be used, and the objective function remains unchanged. The above-mentioned optimized coefficients are the model parameters to be stored.
[0128] At block 640, the computing device 200 calculates and verifies the source term. Figure 3 The process described in block 330 is similar and will not be described in detail. In addition, the computing device 200 can also calculate the source term modulus and source term density distribution based on the model parameters optimized in block 630 to determine the effect of the extraction mechanism. If the source term distribution is approximately a Gaussian distribution with a mean of 0, it means that the mechanism is well extracted.
[0129] In some embodiments, the computing device 200 also determines a method for solving the linear system during decompression based on the template corresponding to the optimized parameters. If the coefficient template is full rank (non-singular matrix), the linear system of equations can be directly solved using one layer of boundary data; if the coefficient template is ill-conditioned (singular matrix), two layers of boundary data are used to solve the overdetermined system through a generalized inverse method.
[0130] In some embodiments, the computing device 200 may also truncate the source item according to the data precision. In the case of lossless compression, the number of source item truncation bits and the corresponding reconstruction residual will change the compression rate, which is affected by the data scenario and data quality. Therefore, a fixed preset number of bits may be used according to the data scenario and test analysis, or further optimized during the compression process to adjust the number of truncation bits to achieve the best effect. In addition, to ensure the effectiveness of compression, other models (such as the Lorenzo prediction model) may be used for verification and comparison.
[0131] In block 650, the computing device 200 encodes and stores boundary elements, source items, and model parameters. Taking the above-mentioned nine-grid template as an example, the data to be stored are the above-mentioned template parameters (8 coefficients), source items (if the original data is a matrix of size N×M, the number of source items is (N-1)×(M-1)), and one or two layers of boundaries. The computing device 200 encodes the above-mentioned data to be stored and outputs compressed data.
[0132] Figure 7FIG. 7 is a schematic flow chart of another process 700 for data decompression according to an embodiment of the present disclosure. Figure 4 An exemplary implementation of process 400 is described. It should be understood that one or more steps of process 700 may be omitted, and there may be more steps.
[0133] In response to receiving the compressed data, the computing device 200 parses and obtains source items, boundary elements, and template parameters in block 710. In some embodiments, for lossless compression, the computing device 200 may also obtain a residual, ie, a precision truncation error.
[0134] At block 720, computing device 200 performs reverse preprocessing. If preprocessing was performed during the generation of the compressed data (eg, block 620), computing device 200 performs reverse preprocessing. For example, the reverse preprocessing operator may be the inverse operator of the preprocessing operator.
[0135] At block 730, the computing device 200 determines the type of the system of equations. The computing device 200 assembles the stiffness matrix of the linear system coefficients according to the difference format template corresponding to the model parameters and the linear system inversion method determined during the compression process. An exemplary stiffness matrix assembly may be as follows: Assume that the data {u ij} are arranged in one-dimensional form as {u I}, that is, the subscript satisfies Mapping relationship; Initialize the stiffness matrix, Assume that u in the differential template ij and adjacent elements u i-1,j+1 ,…u i+1,j-1 The corresponding one-dimensional subscripts are Then the stiffness matrix K I,I =1(u ij is the boundary point). The right-hand term b I =f ij ,(i,j non-boundary point),b I =u ij (u ij For singular cases of coefficient template matrix, matrix K can be increased by 2M+2N-12 rows to match the values of the second layer boundary.
[0136] In block 740, the computing device 200 solves the system of equations. For the case where the template matrix is non-singular, the linear system of equations can be solved directly. For the case where the coefficient template matrix is ill-conditioned, the overdetermined system of equations obtained in the decompression step 730 can be solved by a generalized inverse method. The above stiffness matrix is a banded sparse matrix, so it is processed and solved as a sparse matrix, which has a fast calculation speed and low memory overhead. After obtaining the solution to the equation, the array is rearranged to obtain the original data matrix shape and the residual matrix (precision truncation error) is superimposed to obtain the final decompressed data.
[0137] Experimental results analysis
[0138] The following selects several scientific computing data to illustrate that the compression algorithm has the best performance when the data itself is a differential equation solution. It should be noted that in most cases, although the data satisfies the differential equation, it is not necessarily a two-dimensional second-order linear equation. In this case, the algorithm proposed in the present disclosure still has good applicability.
[0139] Fig. 8A The schematic diagram of data compression for accurate solution of elliptic equation according to an embodiment of the present disclosure is shown. For the elliptic equation example, the analytical solution u(x,y)=xy+e 2x cos 2y, (x, y) ∈ [-1, 1] × [-1, 1], is discretized into 100 × 100 blocks, the data precision is 64-bit double precision, and the data distribution is as follows Fig. 8A (a). It should be pointed out that for larger arrays, on the one hand, they can be divided into blocks and processed in parallel, and the calculation processes between blocks are independent of each other; on the other hand, from the perspective of local linearization, for variable coefficients or even nonlinear equations, linear second-order equations can still be used to approximate the local area well.
[0140] For the original 64-bit double-precision binary file, its size is 76K, and after zip encoding, it is 73K, with a compression rate of 96%. It can be seen that the general compression method cannot effectively compress high-precision mechanism data. After using the Lorenzo predictor, its source term distribution is as follows Fig. 8A (b) is 1.e-2, which is smaller than the original data level O(1). Therefore, it can be predicted that the information entropy will be reduced under the same precision. After the double-precision source item is encoded, the file size is 64K, and the compression rate is 84%. It should be pointed out that the Lorenzo predictor corresponds to a discrete form of hyperbolic equation. Therefore, for the solution of the elliptic equation, it is inevitable that there is a model bias, so its compression effect has room for further improvement. After using this method, the source item distribution is as follows Fig. 8A(c), the source term is at 1.e-7, which is significantly lower than the source term corresponding to the Lorenzo predictor. After the source term is truncated with 1.e-8 precision, the total storage file size of the final source term and residual (lossless requirement) after encoding is 53K, with a compression rate of 70%. It should be noted that despite the use of the solution of the harmonic equation, the differential equation corresponding to the model of this method still has a discrete error with the original differential equation. Therefore, although the source term is very small, there is still storage overhead.
[0141] Figure 8B A schematic diagram of data compression of the numerical solution of the NS equation according to an embodiment of the present disclosure is shown. For the data of the numerical solution of the incompressible fluid Navier-Stokes (NS), since the data value distribution is not sparse, it is difficult for traditional general coding to obtain ideal compression, as shown in the following Table 1
[0142] Table 1: Comparison of scientific computing data compression effects
[0143]
[0144] The method proposed in this disclosure can significantly optimize the compression effect. The equation has three dimensions, time and space, and is a nonlinear equation. For a certain instantaneous stream function distribution data ( Figure 8B ) is compressed, the data points are 101×81, and the data accuracy is 10 -9 . Unlike the above elliptic equation example, the time section data does not correspond to a two-dimensional second-order homogeneous equation, so the use of the nine-square difference template is equivalent to a local approximation of the mechanism. Under the requirement of losslessness, this method can further optimize the template coefficients on the basis of the Lorenzo method to obtain a source term distribution with a smaller modulus, thereby achieving further compression effect.
[0145] It can be seen that the embodiments of the present disclosure can at least solve or improve the following technical problems: First, for high-precision mechanism data, when general compression methods are almost unable to compress, the embodiments of the present disclosure achieve effective compression through the mechanism model compression method, reaching a compression rate of 70% or even lower. Second, the embodiments of the present disclosure have good adaptability to various types of mechanism data, and even if the control equation of the mechanism itself is nonlinear or high-dimensional, the compression effect can be obtained.
[0146] Compared with the prior art, the improvements of the embodiments of the present invention include the following aspects. Compared with the general image compression method, the present method can effectively compress high-precision mechanism data. Mechanism data, especially the data of differential equation solutions, are often not sparse, the variables change continuously, the values are widely distributed, and the corresponding information entropy is high. Therefore, it is difficult to obtain effective compression with traditional image compression or encoding methods. In addition, compared with fixed predictors, the present method can broaden the applicability of data, and can adaptively optimize different data to obtain a model that better fits the data, thereby reducing model deviation and obtaining better compression effects. In addition, when lossy compression is allowed, since the random noise part can be discarded, the present method will further highlight the advantages.
[0147] Actual remote sensing data or other observation data often have a certain level of random noise. Therefore, during the reconstruction process, noise will affect the compression effect. Pure white noise data cannot be effectively compressed using the mechanism model. On the other hand, the presence of noise increases the number of small-scale structures, which becomes interference for the template optimization process and makes it difficult to obtain the optimized template coefficient. In some implementations, preprocessing technology can be used, that is, before optimizing the extraction mechanism, the original data is compressed with a fixed template, such as the following operator
[0148] 0 1 / 5 0 1 / 5 1 / 5 1 / 5 0 1 / 5 0
[0149] The average operator shown is At this time, the smoothed distribution field It has the effect of noise reduction, and the mechanism is further compressed by an optimized method, making it easy to obtain the mechanism. Then the inverse operator of the average operator is applied to finally obtain the original data.
[0150] The specific technical effects are given below using sea surface temperature remote sensing data as an example. Fig.9A shows an exemplary histogram distribution of sea surface temperature remote sensing data before and after compression according to an embodiment of the present disclosure, Fig. 9B An exemplary mechanism extraction effect of sea surface temperature remote sensing data according to an embodiment of the present disclosure is shown.
[0151] It can achieve a similar technical effect as the previous experiment. Fig.9A It can be seen that the original data has a large small-scale structure, which can be considered to have random noise, and its numerical distribution has multiple peaks. Fig. 9B As shown in Table 2, after the above method is used to reduce noise and extract the mechanism, the residual data distribution is approximately normal. It can be seen that this method can effectively extract the data mechanism and remove the random part. The relevant compression effect is shown in Table 2.
[0152] Table 2: Comparison of sea surface temperature data compression effects
[0153] Original data + encoding Mechanism extraction + encoding Compression 14MB 10MB 71%
[0154] In summary, the data compression scheme proposed in this disclosure is based on the idea of learning. For scenarios that are not easy to model directly, have complex physical mechanisms, and have difficult to measure model parameters, the model is learned through data. , Thus, the optimal result under the same order model is obtained. In some embodiments, the type of equation corresponding to the data is judged, a prediction neighborhood model containing a second-order neighborhood is constructed, the coefficients are initialized, and a fuzzy operator is used for preprocessing. Then, the coefficients of the prediction model are calculated by the data to be compressed, and the residual is calculated by prediction reasoning. The residual is regarded as a source term, and the source term, boundary term and coefficient are entropy encoded. Therefore, the embodiment of the present disclosure is superior to the fixed predictor compression algorithm such as Lorenzo. For data with significant intrinsic mechanism, it can even achieve significant compression effect compared with the general non-mechanism model compression algorithm. In addition, some embodiments of the present disclosure are also based on the idea of local linearization and are compatible with the theory of well-posedness of differential equations. Further, in some embodiments, the compression and decompression efficiency can be further improved by adaptively dividing the data into blocks and parallelizing them, and using a fast optimization algorithm. In addition, the model and data quality, such as the model signal-to-noise ratio, can also be evaluated based on the compression results. It should be noted that the method proposed in the present disclosure can also be extended to three-dimensional data or higher-dimensional data, or more complex coefficient templates such as 4×4 point format. Corresponding to higher-order differential equations, a model containing nonlinear combinations of data points can be adopted.
[0155] The data compression and decompression scheme proposed in the present disclosure has the following characteristics: the compression speed is greater than the decompression speed. The main computational effort of the compression process lies in coefficient optimization, in which the computational effort of the objective function is not large and the model parameters are few, and thus has little impact on the compression speed. The decompression process involves solving a system of linear equations, which will further increase the computational effort and memory usage when the compression block increases. Additionally, for a single data block, the overall trend of the compression rate decreases as the data block increases. This is because the local linear assumption can better fit the data mechanism, so appropriate data segmentation is beneficial. Additionally, the compression effect of scientific computing data is better than that of observational data, because the observational data contains more noise that is not reflected in the equation. In addition, the compression effect of equation calculation data is significantly better than that of the method using empirical models.
[0156] Exemplary Apparatus and Implementations
[0157] Fig.10A schematic block diagram of an apparatus 1000 for data compression according to an embodiment of the present disclosure is shown. The apparatus 1000 may be implemented at a computing device 200. The apparatus 1000 includes a model generation unit 1010, a model optimization unit 1020, a model error determination unit 1030, and a data storage unit 1040. The model generation unit 1010 is configured to generate a prediction neighborhood model for multidimensional data to be compressed. The model optimization unit 1020 is configured to optimize the model parameters of the prediction neighborhood model based on the multidimensional data. The model error determination unit 1030 is configured to determine the model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, and the model error is associated with the internal elements of the multidimensional data. The data storage unit 1040 is configured to store the model error, the optimized model parameters, and the boundary elements of the multidimensional data.
[0158] In some embodiments, the model parameters of the prediction neighborhood model may indicate that an element of the multidimensional data is associated with at least one neighboring element of the element. In some embodiments, the multidimensional data may be two-dimensional data, and the prediction neighborhood model may include a template covering the elements in the two-dimensional data. In some embodiments, the multidimensional data includes mechanism data, such as scientific computing data or observational data.
[0159] In some embodiments, the apparatus 1000 may further include a blocking unit, where the blocking unit is configured to obtain multi-dimensional data by blocking the input data, where the multi-dimensional data is a portion of the blocked input data.
[0160] In some embodiments, the apparatus 1000 may further include a truncation unit configured to generate multidimensional data and a precision truncation error that conforms to the data precision from the input data according to the data precision; and store the precision truncation error.
[0161] In some embodiments, the model generation unit 1010 may also be configured to determine a plurality of candidate prediction neighborhood models having different model parameters; determine the model error of the multidimensional data relative to each candidate prediction neighborhood model; and determine the candidate prediction neighborhood model having the minimum model error as the prediction neighborhood model for the multidimensional data. In some embodiments, the plurality of candidate prediction neighborhood models include at least one of the following: an elliptic type difference operator; a hyperbolic type difference operator; or a parabolic type difference operator.
[0162] In some embodiments, the device 1000 may also include a preprocessing unit, which is configured to generate preprocessed multidimensional data from the multidimensional data based on a preprocessing template for noise reduction, and the model optimization unit 1020 may be configured to optimize model parameters of the prediction neighborhood model based on the preprocessed multidimensional data.
[0163] In some embodiments, the model optimization unit 1020 can be configured to perform the following actions at least once: determine the predicted values of at least some elements of the multidimensional data based on the model parameters of the prediction neighborhood model; and update the model parameters by minimizing the difference between the actual values and the predicted values of at least some of the elements.
[0164] In some embodiments, the model error determination unit 1030 can also be configured to: determine the predicted value of the internal element of the multidimensional data based on the optimized model parameters and the boundary elements of the multidimensional data; and determine the model error based on the difference between the actual value and the predicted value of the internal element.
[0165] In some embodiments, the model error determination unit 1030 may also be configured to obtain a truncated model error for storage according to data accuracy.
[0166] In some embodiments, the data storage unit 1040 may also be configured to: determine the boundary elements of the multidimensional data to be stored based on the optimized model parameters; and store the determined boundary elements. In some embodiments, the data storage unit 1040 may also be configured to determine a layer of boundary elements of the multidimensional data for storage in response to the optimized model parameters corresponding to a non-singular matrix template; or determine a layer of boundary elements of the multidimensional data for storage in response to the optimized model parameters corresponding to a singular matrix template.
[0167] In some embodiments, the apparatus 1000 may further include an evaluation unit configured to evaluate the quality of the prediction neighborhood model and the multidimensional data based on the distribution of the model error.
[0168] Fig.11 A schematic block diagram of an apparatus 1100 for data decompression according to an embodiment of the present disclosure is shown. The apparatus 1100 may be implemented at a computing device 200. The apparatus 1100 includes an acquisition unit 1110, an equation group determination unit 1120, and a solution unit 1030. The acquisition unit 1110 is configured to acquire boundary elements of stored multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data. The equation group determination unit 1120 is configured to determine an equation group for the multidimensional data based on the model parameters, the boundary elements, and the model errors. The solution unit 1130 is configured to determine the internal elements of the multidimensional data by solving the equation group.
[0169] In some embodiments, the model parameters of the prediction domain model may indicate that an element of the multidimensional data is associated with at least one neighboring element of the element. In some embodiments, the multidimensional data may be two-dimensional data, and the prediction neighborhood model may include a template covering the elements in the two-dimensional data. In some embodiments, the multidimensional data includes mechanism data, such as scientific computing data or observational data.
[0170] In some embodiments, the apparatus 1100 may further include a precision error unit configured to obtain a precision truncation error of the stored multidimensional data; and accumulate the determined internal elements and boundary elements of the multidimensional data with the corresponding precision truncation error.
[0171] In some embodiments, the system of equations determination unit 1120 may also be configured to: determine the matrix of the system of equations based on the model parameters; and determine the components of the column vector of the system of equations based on the boundary elements and the model errors.
[0172] In some embodiments, the solving unit 1130 may also be configured to: determine a method for solving the system of equations based on the type of the matrix; and solve the system of equations using the determined method. In some embodiments, when the type of the matrix is a non-singular matrix, the boundary elements of the stored multidimensional data include a layer of boundary elements of the multidimensional data, or when the type of the matrix is a singular matrix, the boundary elements of the stored multidimensional data include a second layer of boundary elements of the multidimensional data.
[0173] The present disclosure may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present disclosure.
[0174] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.
[0175] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.
[0176] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may 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 may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.
[0177] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.
[0178] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.
[0179] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.
[0180] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.
[0181] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A method for data compression, comprising: generating a predictive neighborhood model for the multidimensional data to be compressed; Based on the multidimensional data, optimizing model parameters of the prediction neighborhood model; Determining a model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, the model error being associated with internal elements of the multidimensional data; as well as The model errors, the optimized model parameters, and boundary elements of the multidimensional data are stored.
2. The method according to claim 1, wherein: The model parameters of the prediction neighborhood model indicate that an element of the multidimensional data is associated with at least one neighboring element of the element.
3. A method according to claim 1 or 2, wherein the multidimensional data is two-dimensional data, and the prediction neighborhood model includes a template covering elements in the two-dimensional data.
4. The method according to any one of claims 1 to 3, further comprising: The multi-dimensional data is obtained by partitioning the input data into blocks, and the multi-dimensional data is a part of the partitioned input data.
5. The method according to any one of claims 1 to 4, further comprising: According to the data precision, generating the multidimensional data and precision truncation error that meet the data precision from the input data; as well as The precision truncation error is stored.
6. The method according to any one of claims 1 to 5, wherein generating a prediction neighborhood model for the multidimensional data to be compressed comprises: determining a plurality of candidate prediction neighborhood models having different model parameters; determining a model error of the multidimensional data relative to each candidate prediction neighborhood model; as well as A candidate prediction neighborhood model with a minimum model error is determined as the prediction neighborhood model for the multidimensional data.
7. The method according to claim 6, wherein the plurality of candidate prediction neighborhood models comprises at least one of the following: Elliptic type difference operators; a hyperbolic type difference operator; or Parabolic type difference operator.
8. The method according to any one of claims 1 to 7, further comprising: Generating preprocessed multidimensional data from the multidimensional data based on a preprocessing template for noise reduction, and optimizing model parameters of the prediction neighborhood model includes: Based on the preprocessed multidimensional data, the model parameters of the prediction neighborhood model are optimized.
9. The method according to any one of claims 1 to 8, wherein optimizing the model parameters of the prediction neighborhood model comprises performing the following actions at least once: Determining predicted values of at least some elements of the multidimensional data based on the model parameters of the prediction neighborhood model; and The model parameters are updated by minimizing the difference between the actual value and the predicted value of the at least part of the elements.
10. The method according to any one of claims 1 to 9, wherein determining the model error of the prediction neighborhood model comprises: determining predicted values of the internal elements of the multidimensional data based on the optimized model parameters and the boundary elements of the multidimensional data; as well as The model error is determined based on a difference between the actual value and the predicted value of the internal element.
11. The method of claim 10, wherein determining the model error comprises: Based on the data precision, the truncated model error is obtained for storage.
12. The method according to any one of claims 1 to 11, wherein storing the boundary elements of the multidimensional data comprises: Based on the optimized model parameters, determining boundary elements of the multidimensional data to be stored; as well as The determined boundary elements are stored.
13. The method according to claim 12, wherein determining the boundary elements of the multidimensional data to be stored comprises: In response to the optimized model parameters corresponding to a non-singular matrix template, determining a layer of boundary elements of the multidimensional data for storage; or In response to the optimized model parameters corresponding to a singular matrix template, two-level boundary elements of the multi-dimensional data are determined for storage.
14. The method according to any one of claims 1 to 13, further comprising: Based on the distribution of the model errors, the quality of the prediction neighborhood model and the multidimensional data is evaluated.
15. The method according to any one of claims 1 to 14, wherein the multidimensional data comprises mechanistic data.
16. A method for data decompression, comprising: Obtaining stored boundary elements of multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data; determining a system of equations for the multidimensional data based on the model parameters, the elements and the model errors; as well as By solving the system of equations, internal elements of the multidimensional data are determined. 17 . The method of claim 16 , wherein the model parameter of the predictive domain model indicates that an element of the multidimensional data is associated with at least one neighboring element of the element.
18. The method according to claim 16 or 17, wherein the multidimensional data is two-dimensional data, and the prediction neighborhood model includes a template covering elements in the two-dimensional data.
19. The method according to any one of claims 16 to 18, further comprising: Obtaining the precision truncation error of the stored multidimensional data; as well as The determined internal elements and boundary elements of the multidimensional data are accumulated with the corresponding precision truncation errors.
20. The method according to any one of claims 16 to 19, determining a system of equations for internal elements of the multidimensional data comprises: determining a matrix of the system of equations based on the model parameters; as well as Components of a column vector of the system of equations are determined based on the boundary elements and the model errors.
21. The method of any one of claims 16 to 20, wherein solving the system of equations comprises: Determining a method for solving the system of equations based on the type of the matrix; as well as The system of equations is solved using the determined method.
22. The method of claim 21, wherein: In the case where the type of the matrix is a non-singular matrix, the boundary elements of the stored multi-dimensional data include a layer of boundary elements of the multi-dimensional data; or In a case where the type of the matrix is a singular matrix, the boundary elements of the stored multi-dimensional data include two-level boundary elements of the multi-dimensional data.
23. A method according to any one of claims 16 to 22, wherein the multidimensional data comprises mechanistic data.
24. An electronic device comprising: processing unit and memory, The processing unit executes the instructions in the memory so that the electronic device performs a method, the method comprising: generating a predictive neighborhood model for the multidimensional data to be compressed; Based on the multidimensional data, optimizing model parameters of the prediction neighborhood model; determining a model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, the model error being associated with internal elements of the multidimensional data; and The model errors, the optimized model parameters, and boundary elements of the multidimensional data are stored.
25. An electronic device comprising: processing unit and memory, The processing unit executes the instructions in the memory so that the electronic device performs a method, the method comprising: Obtaining stored boundary elements of multidimensional data, model parameters of a prediction neighborhood model, and model errors of the prediction neighborhood model, wherein the model errors are associated with internal elements of the multidimensional data; determining a system of equations for the multidimensional data based on the model parameters, the elements, and the model errors; and By solving the system of equations, internal elements of the multidimensional data are determined.
26. An apparatus for data compression, comprising: A model generation unit, configured to generate a prediction neighborhood model for the multidimensional data to be compressed; A model optimization unit, configured to optimize model parameters of the prediction neighborhood model based on the multidimensional data; a model error determination unit configured to determine a model error of the prediction neighborhood model based on the multidimensional data and the optimized model parameters, wherein the model error is associated with an internal element of the multidimensional data; as well as The data storage unit is configured to store the model error, the optimized model parameter and the boundary elements of the multi-dimensional data.
27. A device for data decompression, comprising: an acquisition unit configured to acquire boundary elements of the stored multidimensional data, model parameters of a prediction neighborhood model, and a model error of the prediction neighborhood model, wherein the model error is associated with internal elements of the multidimensional data; an equation group determination unit configured to determine an equation group for the multidimensional data based on the model parameters, the boundary elements and the model errors; as well as The solving unit is configured to determine the internal elements of the multi-dimensional data by solving the equation group.
28. A computer-readable storage medium having one or more computer instructions stored thereon, wherein the one or more computer instructions are executed by a processor to cause the processor to perform the method according to any one of claims 1 to 15 or any one of claims 16 to 23.
29. A computer program product comprising machine executable instructions which, when executed by a device, cause the device to perform a method according to any one of claims 1 to 15 or any one of claims 16 to 23.
