A multi-modal data augmentation storage system based on parquet format

Abstract

The application provides a multi-modal data enhancement storage system based on a Parquet format, and realizes unified storage and optimization of multi-modal data through native modification and expansion of the Parquet format. The system comprises a modal perception format expansion component and a modal optimization encoding strategy component: the former realizes deep fusion of multi-modal semantics and storage structure by expanding a Parquet logical type system and constructing a four-layer hierarchical metadata architecture from a file level to a page level; and the latter is fused with a Parquet compression algorithm instruction level, performs targeted storage optimization based on a semantic descriptor and modal characteristic parameters, and solves the problems of unified storage and efficient management of multi-modal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data storage, and particularly relates to a multi-modal data enhancement storage system based on a Parquet format. BACKGROUND

[0002] In the tide of rapid development of information technology, the evolution of data storage format is always closely related to the richness of data types and the diversification of application scenarios. With the explosive growth of various types of data, the demand for efficient and flexible data storage solutions is increasingly urgent. Apache Parquet, as a high-efficiency columnar storage format, has rapidly stood out and been widely applied in the field of big data processing due to its excellent compression rate and superior query performance. The columnar storage method enables Parquet to compress and encode data for different columns separately, greatly reducing the occupation of storage space, and in the query process, only the required column data can be read, significantly improving the query efficiency, and providing strong support for structured data processing in the era of big data.

[0003] However, with the rapid development of artificial intelligence and multi-modal analysis technology, data forms are no longer limited to traditional structured data, and a large amount of multi-modal data such as images, audios, videos, and vectors has emerged. The limitations of the traditional Parquet format have become increasingly prominent, and it has been unable to meet the needs of modern data processing systems for unified management of multi-modal data. The traditional Parquet format is mainly designed for structured data, and its type system only supports basic data types such as integers, floating-point numbers, and strings. When faced with unstructured data such as images, audios, and videos, Parquet can only store them as binary objects as a whole, and cannot recognize and utilize the internal structure and characteristics of these data. This processing method brings a series of problems: in terms of storage efficiency, due to the inability to choose the optimal compression strategy for different modal data characteristics, it leads to waste of storage space; in terms of query performance, it is difficult to filter and optimize based on modal-specific attributes, which limits the query speed; at the same time, it lacks the ability to describe metadata for multi-modal data characteristics, making it difficult to support advanced analysis functions and restricting the full exploitation of the value of multi-modal data. Multi-modal data storage has unique technical needs, and different types of data have their own distinct characteristics. Image data needs to consider resolution, color space, and compression format; audio data focuses on sampling rate, bit depth, and encoding method; video data involves frame rate, codec, and key frame distribution; vector data needs to support high dimensionality, different precision, and sparse representation. Traditional storage solutions usually use specialized file formats to handle these different types of data, which results in data being scattered in different storage systems, increasing the complexity of data management, making data integration, migration, and maintenance more difficult, and hindering cross-modal joint analysis, and unable to fully exploit the advantages of multi-modal data in fusion analysis. SUMMARY

[0004] To solve the above technical problems, the technical scheme adopted by the present application is:

[0005] The embodiment of the application provides a multi-modal data enhancement storage system based on a Parquet format, the system realizes unified storage and optimization of multi-modal data through native modification and expansion of the Parquet format, and comprises a modal perception format expansion component and a modal optimization coding strategy component; wherein the modal perception format expansion component realizes deep fusion of multi-modal data semantics and a Parquet storage structure by performing targeted expansion on a Parquet format logical type system and a metadata architecture, thereby establishing a semantic description system of multi-modal data, and comprises a logical type system expansion sub-module and a metadata structure enhancement sub-module; wherein the logical type system expansion sub-module is realized by expanding a logical type definition of the Parquet, inherits a LogicalType base class of the Parquet and adds a modal specific parameter field; the metadata structure enhancement sub-module is used for constructing a four-layer metadata architecture at a file level, a row group level, a column level and a page level, the architecture is accurately mapped with a page structure of Parquet physical storage, is used for recording global distribution, local features and fine positioning information of multi-modal data in layers, and realizes deep coupling of modal feature information and the Parquet storage structure; the modal optimization coding strategy component realizes instruction level fusion with a Parquet compression algorithm, is used for receiving a semantic descriptor generated by the modal perception format expansion component, and performs targeted storage optimization based on type characteristics, feature parameters and optimization suggestions of multi-modal data.

[0006] The multi-modal data enhancement storage system based on the Parquet format provided by the embodiment of the application realizes three core technical breakthroughs by modifying the Parquet native format. Firstly, deep coupling of a logical type system and physical storage, unlike a traditional scheme of externally connecting metadata, the application directly embeds modal semantics in a logical type definition of the Parquet, so that the storage system can natively understand multi-modal data characteristics on the basis of columnar storage. Secondly, alignment of a four-layer metadata architecture and a page structure of columnar storage, through hierarchical metadata organization at a file level, a row group level, a column level and a page level, accurate mapping with the Parquet physical storage structure is realized, and the efficiency of metadata access and the accuracy of data positioning are ensured. Finally, instruction level fusion of a modal optimization coding strategy and a Parquet compression algorithm, by identifying coding characteristics of compressed media data, intelligent skipping and native coding reservation are realized in the compression pipeline of the Parquet, and performance loss caused by secondary compression in the traditional scheme is avoided. The application modifies the Parquet format systematically, retains the advantages of efficient processing of structured data while expanding the support capability for multi-modal data. By establishing a unified storage model and modal perception technology, more efficient bottom technology support is provided for multi-modal data analysis and artificial intelligence application.

[0007] It is to be understood that the embodiments described herein are merely exemplary of the application and that a myriad of modifications, both as to the nature and number of elements within the execution of the application and as to the modes of execution thereof, can be made by those skilled in the art, without exceeding the scope of the application as defined by the appended claims. Accordingly, the particular embodiment discussed is illustrative only and not restrictive. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0009] Figure 1 A structural block diagram of a multi-modal data enhancement storage system based on a Parquet format provided by the embodiments of the present application. DETAILED DESCRIPTION

[0010] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this description, the singular forms "a", "an" and "the" include plural references unless the context clearly dictates otherwise. The term "and / or" includes any and all combinations of one or more of the associated listed items.

[0012] It should be noted that some of the example embodiments described are depicted as processes or methods that are depicted as flowcharts. Although the processes are described in a particular, sequential order, many of the processes described can be performed concurrently, in parallel, or simultaneously. In addition, the order of individual processes can be re-arranged. The processes can terminate when their operations are completed, but can also terminate without completing their operations due to "failure" or upon cancellation for other reasons. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0013] The embodiment of the application provides a multi-modal data enhancement storage system based on a Parquet format, which realizes unified storage and optimization of multi-modal data by native modification and expansion of the Parquet format, so as to establish a unified storage model and a modal perception technology of multi-modal data. The unified storage model refers to simultaneously storing and managing data of different modalities such as images, audios, videos and vectors in a single Parquet file, and realizing cross-modal data organization and access through a unified logical type system and a metadata architecture. The modal perception technology refers to that the storage system can recognize, understand and utilize the inherent characteristics of different modal data such as the resolution of images, the sampling rate of audios, the GOP structure of videos and the sparsity of vectors, and perform targeted storage optimization and query acceleration based on these characteristics.

[0014] In the embodiment of the application, as shown in Figure 1 The system provided by the embodiment of the application can include a modal perception format expansion component 1 and a modal optimization encoding strategy component 2, and the Parquet format is modified through the two components. The two core technical components cooperate with each other to form a complete technical chain from data understanding to storage optimization. The modal perception format expansion component enhances the logical type system and the metadata structure to enable the Parquet format to understand and describe the characteristics of multi-modal data; and the modal optimization encoding strategy is based on the understanding ability to realize targeted storage optimization and performance improvement.

[0015] In order to quantitatively evaluate the improvement effect of the application, a standardized mathematical evaluation model needs to be established. The establishment of the model is based on the following considerations: the storage and query optimization effect of multi-modal data needs a unified quantitative standard; the characteristic differences of different modal data require a hierarchical evaluation mechanism; the weight requirements of business scenarios for storage cost and query performance are different, and a flexible weight configuration mechanism is needed.

[0016] Let the original multi-modal data set be D={d1, d2, …, d i , …, d n}, wherein the i-th data item d i has a determined modal type m i ∈{IMAGE, AUDIO, VIDEO, VECTOR}, the value of i is 1 to n, n is the number of data items in D, IMAGE is an image type, AUDIO is an audio type, VIDEO is a video type, and VECTOR is a vector type.

[0017] In order to avoid the problems of non-uniform dimension and logical contradiction in calculation, the application adopts a hierarchical evaluation model, first defines a basic efficiency index at the data item level, then integrates it into a unified technical benefit index through a business weight mechanism, and finally aggregates it to the data set level.

[0018] For each data item d i , two basic efficiency indicators are defined: normalized storage efficiency σ i and normalized query efficiency improvement T i , and then based on these two indicators to build the data item d i transformation effect representation value G i .

[0019] Wherein, the normalized storage efficiency σ i is defined as: σ i =max (0, 1-S i / C i baseline ).

[0020] Wherein, S i represents the storage size after the data item d i is transformed, C i baseline is the benchmark size when the data item d i is stored as a binary object by the traditional Parquet. C i baseline is obtained through a pre-established standard test data set, which contains representative samples of each modality: image type accounts for 30% (including JPEG, PNG, TIFF format, resolution from 720p to 4K), audio type accounts for 25% (including MP3, AAC, WAV format, sampling rate from 16kHz to 192kHz), video type accounts for 25% (including H.264, H.265, VP9 encoding, resolution from 1080p to 4K), vector type accounts for 20% (dimension from 100 to 10000, sparsity from 0.1 to 0.9). The size of all samples stored using the traditional Parquet BINARY type and applying the Snappy default compression algorithm is used as the benchmark value.

[0021] When S i <C i baseline , σ i >0, indicating that storage space is saved, the larger the value, the more significant the saving effect; when S i =C i baseline , σ i =0, indicating that the storage size is the same as the benchmark, no saving effect; when S i >C i baseline , σ i =0 is constrained by the max function to avoid negative values affecting comprehensive evaluation caused by storage expansion. Obviously, σ iThe value range of the storage efficiency is [0, 1], wherein 1 represents a theoretically complete compression (actually unattainable), and 0 represents no compression effect or storage expansion. The physical meaning of the storage efficiency is the space saving degree of the quantitative reconstruction scheme relative to the traditional scheme, which provides a direct numerical basis for storage cost optimization.

[0022] The normalized query efficiency T i is defined as: T i = (γ(m i )-1) / (γ(m i )+1) x H(γ(m i )-1). Wherein γ(m i ) is the query performance improvement coefficient of mode m i , γ(m i )=T baseline (m i ) / T new (m i ), T baseline (m i ) is the query execution time of the traditional Parquet processing mode m i data, T new (m i ) is the query execution time of the system after the reconstruction processing mode m i data. H(γ(m i )-1) is the Heaviside step function, if (γ(m i )-1) < 0, H(γ(m i )-1)=0, if (γ(m i )-1) ≥ 0, H(γ(m i )-1)=1. (γ(m i )-1) / (γ(m i )+1) maps the performance improvement coefficient γ(m i ) to the interval (−1, 1), combined with the truncation effect of the Heaviside step function H(γ(m i )-1), so that the final T i is mapped to the interval [0, 1), realizing the dimensional unity with the storage efficiency index; when γ(m i )=1 (no change in performance), T i =0; when γ(m i )>1 (performance improvement), T i >0, and monotonically increases with the increase of performance improvement, but is always less than 1; when γ(m i )<1 (performance decline), through the Heaviside step function, T i =0, ensuring that the performance decline does not negatively contribute to the comprehensive evaluation; when the query fails, it is agreed that Tnew (m i i =+∞, this time γ(m i )=0, T i =0. The γ(m i )+1 ensures the mathematical stability of the fraction in all cases, avoiding the problem of division by zero. The physical meaning of the query efficiency T i is that the degree of time performance improvement of the transformation scheme relative to the traditional scheme is quantified, providing a standardized evaluation benchmark for query performance optimization.

[0023] In actual business scenarios, there is often a trade-off between storage optimization and query optimization, and different application scenarios also have different emphasis on the two. Therefore, a comprehensive evaluation mechanism needs to be established to combine the storage efficiency and query efficiency according to the business requirements.

[0024] The business weight coefficient is defined as follows: α is the business weight coefficient of storage efficiency, reflecting the business's emphasis on storage cost savings; β is the business weight coefficient of query efficiency, reflecting the business's emphasis on query performance improvement; δ is the synergy coefficient, reflecting the importance of the synergy between storage and query efficiency.

[0025] The transformation effect representation value G i is defined as: G i =α×σ i +β×T i +δ×σ i ×T.

[0026] δ is the synergy coefficient, δ∈[0, 0.1] is the synergy coefficient, and the default value is 0.05. The mathematical properties and business implications of this index are reflected in the following aspects: linear weighted combination ensures that the contribution of the two basic indexes is proportional to the business weight, and the synergy term δ×σ i ×T i embodies the synergy between storage efficiency and query efficiency, and when both are high, the overall system benefit will have an additional boost. The constraint condition is α+β+δ≤1 and α, β, δ≥0, which ensures that G i ∈[0,1]. When α=1, β=0, δ=0, G i =σ i , focusing entirely on storage efficiency; when α=0, β=1, δ=0, G i =T i, fully focus on query efficiency; when δ = 0, it degenerates to the original linear combination. The monotonicity of this index ensures that when any base index improves, the comprehensive benefit will not decrease, which meets the intuition of the optimization goal. The physical meaning of the reconstruction effect representation value lies in the unification of multi-dimensional technical improvement effects into a single numerical index, providing a standardized measurement benchmark for the overall evaluation and comparison of technical solutions. At the same time, through the weight mechanism, it adapts to the optimization focus of different business scenarios.

[0027] After establishing the reconstruction effect representation value G i i at the data item level, further consideration of evaluation needs in practical applications is required. Although G i i can accurately reflect the optimization effect of a single data item, in actual technical solution evaluation and system performance analysis, a macro index that reflects the overall performance of the entire data set is still needed.

[0028] This conversion from micro to macro is based on the following practical considerations: multi-modal data sets usually contain millions of data items, and directly comparing technical solutions and monitoring performance based on massive G i i values is neither realistic nor efficient; system administrators and decision-makers need a simple overall performance index to evaluate the overall effect of technical solutions; different sizes of data items have different impacts on system performance, and this difference needs to be reflected in the aggregation process. Therefore, an aggregation mechanism from micro to macro is needed to effectively aggregate technical benefits at the data item level into overall evaluation indexes at the data set level.

[0029] The reconstruction effect representation value G(D) at the data set level is defined as: G(D) = (∑ n i=1 ln(1 + C i baseline ) × G i ) / ∑ n i=1 ln(1 + C i baseline ).

[0030] This formula uses a logarithmic weight based on the reference size of the data item, ln(1 + C i baseline ) as the weight factor, larger data items have a more significant impact on system performance and should have a greater weight in the aggregation evaluation. The logarithmic function ensures the marginal decreasing effect of weight growth, avoiding the excessive dominance of large files in the evaluation results. When all data item sizes are equal, this formula degenerates to simple averaging. This aggregation mechanism realizes the effective conversion from micro evaluation at the data item level to macro evaluation at the data set level, providing a unified quantitative standard for the overall evaluation and comparison of technical solutions.

[0031] The technical scheme solves the problem of global dependence through data item level efficiency calculation and benchmark comparison processing, and ensures the practical availability of the model in the streaming processing environment. Through fractional transformation of the query efficiency index, the storage efficiency and the query efficiency have the same value interval [0, 1], and a reasonable correlation between the two is established, which provides an explicit optimization goal and evaluation standard for the specific implementation of the subsequent technical scheme.

[0032] In practical applications, G i The two indexes G and G(D) bear different levels of evaluation responsibilities: the data item level transformation effect representation value G i is mainly used for real-time encoding strategy decision-making, and selects the most suitable storage optimization scheme for each data item during data writing, ensuring that the system can make immediate strategy adjustment and parameter configuration based on accurate modal feature analysis. The data set level transformation effect representation value G(D) is mainly used for offline performance evaluation and technical scheme verification, and its calculation occasions include overall effect evaluation after batch data processing, regular system performance monitoring (recommended period is daily or weekly), comparison analysis before and after technical scheme deployment, etc. The core use of G(D) is reflected in four aspects: technical scheme verification, verifying the overall technical advantage of the transformation scheme compared with the traditional Parquet format; system performance monitoring, regularly evaluating the overall system running status and performance trend; parameter tuning guidance, dynamically adjusting the configuration of business weight coefficients based on the historical trend of G(D); business decision support, providing standardized quantitative basis for technical investment evaluation and resource allocation decision-making. This layered evaluation mechanism not only guarantees the efficiency of real-time processing, but also provides the integrity of macro evaluation, laying a solid mathematical foundation for the overall optimization and continuous improvement of the technical scheme.

[0033] The traditional Parquet format only supports basic data types, and processes multi-modal data as opaque binary objects, which cannot take advantage of its internal structural characteristics for storage and query optimization, which directly affects the improvement potential of storage efficiency and query efficiency in the aforementioned mathematical model. In order to fully play the guiding role of the mathematical model, the modal perception format expansion component 1 expands the logical type system and metadata architecture of Parquet format in a targeted manner, realizes the deep integration of multi-modal data semantics and Parquet storage structure, and establishes a semantic description system for multi-modal data, so that the storage system can understand the internal structure and characteristics of the data, thereby providing a technical basis for achieving a high transformation effect representation value.

[0034] The modal perception format expansion component 1 can include a logical type system expansion submodule 11 and a metadata structure enhancement submodule 12.

[0035] The logical type system extension submodule 11 realizes deep understanding of data semantics by defining a special multimodal logical type system. Specifically, it realizes the extension of the logical type definition of Parquet, inherits the original LogicalType base class of Parquet, and adds modal specific parameter fields. The logical type system extension submodule 11 establishes four main types: IMAGE, AUDIO, VIDEO, and VECTOR. Each type contains a list of parameters for describing the key characteristics of the modal data. The selection mechanism of the system makes multi-level decisions based on these parameter structures: first, the modal type is identified through file header magic number, sampling rate, array dimension, etc., then the corresponding format parameters, resolution parameters, encoding parameters, etc. are extracted based on the identification result, and finally the subsequent storage strategy selection and query optimization decisions are driven based on these feature parameters. This semantic understanding capability directly serves the optimization goal of the mathematical model: by accurately identifying data characteristics, the system selects the most suitable storage strategy to maximize storage efficiency, including selecting native encoding preservation strategy based on JPEG quality factor, selecting scene boundary blocking strategy based on video GOP structure, and selecting CSR storage format based on vector sparsity, etc.; at the same time, an efficient index structure is established to optimize query efficiency, including establishing a four-layer metadata index architecture based on modal parameters, implementing O(1) time complexity data positioning through OffsetIndex, and dynamically selecting multi-dimensional index or simple index strategy according to query efficiency threshold.

[0036] The IMAGE type defines a complete set of image characteristic parameters, including format parameters supporting identification and processing of five mainstream formats: JPEG, PNG, WebP, TIFF, and BMP; resolution parameters recording width, height, and pixel density information to support size-based query filtering, supporting three color space standards (compliant with ISO / IEC61966-2-1 standard): sRGB, AdobeRGB, and DCI-P3, and calculating the color gamut range through the CIE1976Lab* color space model for color-related analysis and processing; compression parameters record compression algorithm, quality level, and encoding characteristics to guide the selection of storage strategies, including: when the JPEG quality factor is greater than a preset value, for example, 75 (determined based on the visual quality threshold of PSNR>30dB), and the file size ratio is less than a preset ratio, for example, 0.3 (determined based on the compression efficiency statistical analysis of 1000 samples), the native encoding preservation strategy is selected to skip secondary compression, when the image size is greater than a preset resolution, for example, 4K resolution (3840x2160 pixels), the pyramid blocking strategy is selected to support multi-resolution access, and when the pixel density variation coefficient is greater than a preset pixel density coefficient, for example, 0.3 (determined based on gradient amplitude statistics), the content-based adaptive blocking strategy is selected to optimize storage locality.

[0037] The AUDIO type contains a set of acoustic characteristic parameters, in which the sampling rate parameter defines the time resolution of the audio to support quality-related queries and processing, the bit depth parameter describes the amplitude accuracy for audio quality assessment and compression strategy selection, the channel configuration parameter records the mono, stereo, 5.1 surround, 7.1 surround four channel layout information to support spatial audio processing, the encoding parameter contains MP3, AAC, PCM, FLAC, OGG five format information for native encoding reservation and conversion optimization, the specific guidance mechanism includes: when the MP3 bitrate is greater than the preset bitrate, for example, 128 kbps (determined based on the "acceptable quality" threshold in ITU-R BS.1116-3 subjective quality evaluation standard), and the standard MPEG-1 Audio Layer III encoding is adopted, the native encoding reservation strategy is selected to directly reserve the frame structure, when the audio is lossless FLAC encoding, its compression structure is reserved to avoid quality loss, when a silent segment (energy less than -40 dB and duration greater than a preset time, for example, 2 seconds), a music paragraph or a natural boundary of a speech segment is detected, a block strategy based on audio feature change is selected to optimize access efficiency.

[0038] The VIDEO type defines a set of space-time characteristic parameters, in which the resolution parameter contains width, height and pixel ratio information to support video quality grading and query filtering, the frame rate parameter describes the time sampling frequency for playback performance optimization and time-related queries, the codec parameter records H.264, H.265, VP9, AV1 four encoding information to guide storage and decoding strategy, the code rate parameter indicates the video data stream density for bandwidth and storage optimization, the key frame distribution parameter records the organization mode of I frame, P frame, B frame to support video segmentation and fast access, the specific guidance mechanism includes: when the H.264 code rate control mode is constant quality (CRF mode) and the GOP structure is complete, the native encoding reservation strategy is selected to reserve the encoding format, when the scene change boundary is detected (determined based on the FFmpeg scene_cut threshold of 0.3), the scene block strategy based on the GOP structure is selected to ensure that each block contains a complete semantic unit, when the video adopts a double-threshold scene segmentation, accurate semantic boundary recognition is performed through FFmpeg scene_cut combined with HSV histogram verification (Bhattacharyya distance greater than 0.5).

[0039] The VECTOR type defines a mathematical characteristic parameter set, wherein the dimension parameter records the dimension of the vector to support the optimized storage and query of high-dimensional data, the element type parameter describes the basic type such as FLOAT, DOUBLE, INT for precision control and compression selection, the normalization parameter indicates whether to apply the L1, L2, etc. normalization method to support the optimization of vector similarity calculation, and the sparsity parameter records the proportion and distribution characteristics of non-zero elements to enable sparse storage optimization. The specific guidance mechanism includes: when the vector dimension is greater than a first vector dimension, for example, 1000, and the sparsity is less than a first sparse value, for example, 0.1, a dense storage format is selected and a quantization compression technology is applied; when the sparsity is greater than a second sparse value, for example, 0.7, a CSR or CSC sparse format is selected to store only the non-zero elements and their indexes; when the precision requirement allows approximate calculation, a dimension reduction storage is selected to reduce the storage space through principal component analysis or random projection; and when the vector is based on a similarity clustering result, a clustering blocking strategy is selected to organize similar vectors in the same row group to improve query efficiency.

[0040] In the embodiment of the present application, the precision requirement allowing approximate calculation means that after comprehensively evaluating the business scenario demand, data application target and technical feasibility, it is determined that the precision loss caused by approximate calculation is within the preset business acceptable threshold system, and the negative impact caused by the loss is less than the benefits generated by approximate processing such as storage efficiency improvement and access speed optimization. The specific determination criteria include:

[0041] Business scenario fault tolerance adaptation: for non-precise matching scenarios such as similarity search, trend prediction, clustering analysis, etc., the core goal is to obtain an effective approximate result that meets the business demand, rather than a unique exact solution. A small precision loss such as a vector similarity calculation error ≤5% and a classification accuracy drop ≤2% will not affect the effectiveness of business decision-making.

[0042] Quantitative precision index controllable: the precision loss is quantified by a preset technical threshold (such as a reconstruction error threshold ε and a feature retention threshold), and when the data after approximate calculation is verified, the reconstruction error (such as L2 norm error) ≤ε and the core feature retention (such as variance contribution) ≥90%, it is determined that the precision loss is controllable.

[0043] Benefit-cost balance verification: the benefits such as storage resource saving (such as compression rate improvement ≥30%) and computing efficiency optimization (such as search delay reduction ≥40%) brought by approximate calculation are significantly higher than the potential cost (such as business adjustment cost and result correction cost) that may be caused by precision loss.

[0044] Further, as shown in Figure 1 The logic type system extension submodule 11 can specifically include:

[0045] A modal type identification unit is configured to identify modal types of data items in multi-modal data and map to extended multi-modal logical types.

[0046] Specifically, the modal type identification unit performs type feature matching and verification on the data items of the input multi-modal data by using a preset modal type identification library (including core type identifiers such as IMAGE_TYPE, AUDIO_TYPE, VIDEO_TYPE, and VECTOR_TYPE), accurately identifies the modal types of the data items, and maps the identification results to an extended multi-modal logical type system. The system inherits the original LogicalType base class of Parquet and establishes a one-to-one correspondence between the modal types and the logical types, thereby providing a basic type basis for subsequent feature extraction and strategy optimization.

[0047] A feature parameter extraction unit is configured to extract feature parameters of corresponding modal types based on the identified modal types and associate the extracted feature parameters to semantic description fields of the logical types.

[0048] Specifically, the feature parameter extraction unit determines the modal types based on the modal type identification unit, calls a feature extraction interface of the corresponding modal type, extracts parameters such as resolution, pixel density distribution, and quality factor for image data, extracts parameters such as bit rate, sampling rate, and spectral feature variation coefficient for audio data, extracts parameters such as code rate control mode, GOP structure, and interframe difference for video data, and extracts parameters such as dimension scale, sparsity distribution, and cosine similarity for vector data. The extracted parameters are associated to the semantic description fields of the logical types in a preset format to form a structured feature parameter set.

[0049] An optimization suggestion generation unit is configured to generate storage strategy optimization suggestions based on the feature parameters and form standardized semantic descriptors containing modal types, feature parameters, and optimization suggestions.

[0050] Specifically, the optimization suggestion generation unit analyzes the feature parameter set output by the feature parameter extraction unit in combination with a preset parameter threshold rule and an optimization strategy library: generates a “native encoding reservation” suggestion when the image quality factor is greater than a preset threshold; generates a “quantization compression + dense storage” suggestion when the vector dimension is greater than a preset dimension and the sparsity is less than a preset value; generates a “semantic boundary block priority” suggestion when the video GOP structure is complete, and the like. Finally, the modal type identifier, the structured feature parameters, and the targeted optimization suggestions are integrated to form standardized semantic descriptors, thereby providing a unified input basis for the modal optimization encoding strategy component.

[0051] The metadata structure enhancement module 12 is used to build a four-layer hierarchical metadata architecture of file level, row group level, column level and page level, which is accurately mapped with the page structure of Parquet physical storage, is used for recording the global distribution, local features and fine positioning information of multi-modal data, and realizes deep coupling of modal feature information and Parquet storage structure.

[0052] In the embodiment of the application, the four-layer metadata architecture adopts a top-down organization form, follows the design concept of "global overview-local details", and includes a file level metadata storage unit, a row group level metadata storage unit, a column level metadata storage unit and a page level metadata storage unit, which are accurately mapped with the Parquet physical storage structure.

[0053] The file level metadata storage unit is used to extend the FileMetaData structure of Parquet, and provides global view and cross-modal statistical information, such as the distribution of modal types contained in the file, the number and total size of each modal data, etc.; the row group level metadata is used to extend the RowGroup structure of Parquet, and describes the data block organization and modal distribution characteristics, and contains the statistical characteristics and block information of each modal data in the row group; the column level metadata is used to extend the ColumnChunk.meta_data field of Parquet, and provides specific information of column dimension and modal characteristic description, and contains specific numerical values of each modal characteristic parameter set, encoding method identifier and encoding parameter configuration; the page level metadata realizes hybrid storage by using the PageHeader extension area and the data page tail reserved area of Parquet, stores the positioning information of data in the page, and contains the feature summary, semantic label and index positioning identifier of data in the page.

[0054] The hierarchical structure enables the query engine to selectively access metadata of appropriate granularity according to query requirements, realizes multi-level optimization from coarse-grained filtering to fine-grained positioning, and specifically includes that the file level metadata supports file level filtering based on modal type distribution, the row group level metadata supports data skipping and selective reading based on modal correlation, the column level metadata supports predicate pushdown and column pruning, and the page level metadata supports the finest-grained selective access and conditional filtering, thereby avoiding unnecessary metadata processing overhead.

[0055] In order to maintain backward compatibility, the system adopts a two-layer mapping mechanism: in the physical storage layer, the modal data is still stored in the basic column of BINARY or BYTE_ARRAY type, so that the old version reader can normally access the data; and in the logical representation layer, the semantic enhancement is performed on these basic types through the newly added logical type definition, so that the new version reader can understand and utilize the modal characteristics.

[0056] The specific implementation is to add a logical_type field extension in the schema definition of Parquet, identify the specific modal type (such as "IMAGE_TYPE", "AUDIO_TYPE", etc.) through type_name, record modal-specific parameter information through type_parameters, record format version information (such as "MAFE_v1.0" indicating modal awareness format extension version 1.0) through compatibility_version field, and identify compatibility requirements through compatibility_flags bit. Bit 0 indicates whether modal awareness support is required, bit 1 indicates whether optimized encoding support is required, and bits 2-7 are reserved for future extension.

[0057] This implementation provides a complete technical basis for the system selection mechanism: modal type identification is implemented through the type_name field, parameter-driven strategy selection is supported through the type_parameters field, and compatibility selection is controlled through the compatibility_flags field. The modal awareness format extension promotes the semantics of unstructured data to the format level, and through the cooperation of the logical type system and the metadata structure, the storage system can identify the specific modal type and key characteristics of the data, understand the structure organization and access mode of the data, and provide a complete semantic basis for the multi-level judgment structure, native encoding reservation decision, boundary block strategy selection and vector optimization decision in the subsequent modal optimized encoding strategy component.

[0058] The specific field definitions of the four-layer metadata architecture are shown in Table 1:

[0059] Table 1

[0060]

[0061] In the embodiment of the application, the binary storage of the metadata field adopts a hybrid storage strategy to ensure fast random access. The page-level metadata storage rules include: fixed-length fields such as quant_bits, nal_unit_fragmented are stored in the PageHeader extension area, and fixed offset design is adopted to support O(1) time complexity field positioning; variable-length fields such as emulation_prevention_bytes, feature_vector are stored in the reserved area at the end of the data page, and Length-Prefixed encoding is adopted, the first 4 bytes store the data length, and the subsequent bytes store the actual data content, wherein each metadata page contains a 32-byte fixed header, which stores the offset table of the key fields, and the field values are stored in little-endian order, consistent with the Parquet format specification v2.9.

[0062] In the embodiment of the present application, the logical type system extension submodule and the metadata structure enhancement submodule realize deep cooperation through a standardized information transmission mechanism, forming a complete selection decision chain. The modal type and parameter structure recognized by the logical type system extension submodule provide a semantic basis for the metadata structure enhancement submodule, ensuring that the four-layer metadata architecture can accurately record and organize modal feature information. In turn, the hierarchical information management system established by the metadata structure enhancement submodule provides an efficient storage and access mechanism for the parameter application of the logical type system extension submodule. This cooperative mechanism not only provides complete semantic descriptors for subsequent modal optimization encoding strategy components, but more importantly, establishes a technical conversion basis from data understanding to storage optimization, ensuring that each selection decision is based on accurate modal feature analysis. Further, the modal optimization encoding strategy component 2 realizes instruction-level fusion with the Parquet compression algorithm, which is used to receive the semantic descriptors generated by the modal perception format expansion component, based on the type characteristics, feature parameters, and optimization suggestions of multi-modal data, to perform targeted storage optimization.

[0063] Further, as shown in Figure 1 The modal optimization encoding strategy component 2 includes:

[0064] The native encoding reservation submodule 21 is used to reserve the native encoding of the compressed media data to avoid secondary compression.

[0065] The modal feature-based boundary partitioning submodule 22 is used to implement semantic boundary partitioning corresponding to the Parquet row group structure based on modal features.

[0066] The vector data optimization submodule 23 is used to adopt a sparse format or quantization compression that adapts to Parquet columnar storage for vector data.

[0067] Further, the modal optimization encoding strategy component 2 processes the semantic descriptors through a three-level decision mechanism to realize accurate selection and configuration of the encoding strategy. The first-level decision mechanism is used to determine the applicable encoding strategy category based on the modal type identifier and the data integrity verification result; the second-level decision mechanism is used to finely adjust the strategy parameters based on specific values in the modal feature parameter set; and the third-level decision mechanism is used to make a final strategy confirmation and execution plan based on the expected performance effect evaluation result, wherein the expected performance effect evaluation result is determined based on the transformation effect representation value of the data item.

[0068] Further, the first-level decision mechanism specifically realizes the determination of the encoding strategy category by the following steps: first, extracting the modality type identifier in the semantic descriptor generated by the modality-aware format extension component to explicitly determine the modality category of the current data item; second, performing integrity checking on the data item to determine whether the data integrity meets the high confidence standard, i.e., the integrity is greater than 95% through the preset checking rule; finally, based on the cooperative judgment of the modality type and the high confidence integrity result, the basic category of the encoding strategy suitable for the data item is determined. Among them, different modality types correspond to their exclusive basic encoding direction, and the high confidence integrity result ensures that the encoding strategy category can be reliably applied, providing prerequisite support for the fine adjustment of the subsequent decision mechanism. For IMAGE_TYPE data, the basic encoding direction focuses on the preservation of the original encoding structure (such as DCT encoding structure adaptation); for AUDIO_TYPE data, it focuses on frame structure integrity optimization and bit rate adaptation; for VIDEO_TYPE data, it focuses on scene partitioning and code rate control mode matching; for VECTOR_TYPE data, it focuses on sparsity adaptation and dimension optimization direction. When the data integrity checking result reaches high confidence (integrity greater than 95%), it indicates that there is no key information loss and no format damage, which can ensure that the above-mentioned basic encoding strategy based on modality type can be stably applied to the data item, avoiding the failure or degradation of the encoding strategy due to data defects. For data with confidence less than or equal to 95%, the modality optimization encoding strategy can be skipped, and the traditional Parquet BINARY / BYTE_ARRAY storage method can be directly used, while the data integrity state is marked in the compatibility_flags field, and the storage efficiency and query efficiency values of the data item are set to 0 to ensure system stability and performance evaluation accuracy.

[0069] The preset checking rule is a standardized checking mechanism for determining whether the data item integrity meets the high confidence standard, specifically including: data structure integrity checking (verifying the logical continuity of data header, metadata identifier, and data body), key field integrity checking (confirming that the modality type identifier, core feature parameters, and other required fields are not missing), format compliance checking (detecting the matching of data encoding format and modality type, without format disorder or damage), and finally outputting the quantitative evaluation value of data integrity through the comprehensive judgment of the above multi-dimensional checking results.

[0070] Further, the second-level decision mechanism specifically includes:

[0071] For the native encoding preservation strategy, the existing compression state indicator in the modality feature parameters is used for fine judgment, and when the compression ratio exceeds the preset threshold and the quality loss evaluation is within the acceptable range, it is determined that the data item is suitable for native encoding preservation;

[0072] For the boundary partitioning strategy based on modal characteristics, the structured information in the modal characteristic parameters and the access prediction model based on historical query patterns are used to determine the partitioning boundary.

[0073] For the vector data optimization strategy, the evaluation matrix composed of dimension scale, sparsity distribution, precision requirement, data distribution characteristics and storage budget constraints is used for strategy selection.

[0074] Further, for the native encoding reservation strategy, the existing compression state indicator in the modal characteristic parameters is used for fine judgment. When the compression ratio exceeds the preset threshold and the quality loss evaluation is within the acceptable range, it is determined that the data item is suitable for native encoding reservation, which specifically includes:

[0075] For image data, based on its quality-related parameters and file size and original data proportion feature parameters, it is determined whether to retain its native encoding structure;

[0076] For video data, based on its code rate control mode and encoding structure integrity feature parameters, it is determined whether to retain its native encoding format;

[0077] For audio data, based on its bit rate parameter and encoding standard compliance feature, it is determined whether to retain its native frame structure.

[0078] Specifically, for JPEG images, when the quality factor is greater than a preset value, for example, 75, and the file size to original data ratio is less than a preset proportion, for example, 0.3, the system selects to retain its DCT encoding structure; for H.264 videos, when the code rate control mode is constant quality and the GOP structure is complete, the system retains its native encoding format; for MP3 audio, when the bit rate is greater than a preset bit rate, for example, 128 kbps, and the standard MPEG encoding is used, the system directly retains its frame structure.

[0079] For the boundary partitioning strategy based on modal characteristics, the structured information in the modal characteristic parameters and the access prediction model based on historical query patterns are used to determine the partitioning boundary, which specifically includes:

[0080] For image data, based on its size feature parameters and pixel density variation feature parameters, the corresponding partitioning strategy is selected;

[0081] For video data, a double-threshold scene segmentation strategy is used to realize semantic boundary recognition, which includes: based on scene switching detection for preliminary segmentation, combined with inter-frame feature difference degree for secondary verification to confirm boundary effectiveness, and dynamic merging for short segments to ensure the semantic integrity of partitioning;

[0082] For audio data, based on the spectral feature variation parameter and the silent duration feature parameter, the partitioning boundary is set;

[0083] For vector data, based on the clustering results and the dimension distribution characteristics, the subblocks are organized in combination with the vector similarity characteristic parameters, and the balance of the subblock size is ensured through dynamic adjustment.

[0084] Specifically, for image data, when the image size exceeds a preset resolution, such as 4K resolution, a pyramid subblock is adopted, and when the pixel density variation coefficient is greater than a preset pixel density variation coefficient threshold, such as 0.3, a content-based adaptive subblock is adopted.

[0085] For video data, a double-threshold scene segmentation strategy is adopted to realize accurate semantic boundary recognition, and the double-threshold scene segmentation strategy includes:

[0086] (1) Preliminary segmentation based on FFmpeg scene_cut detection, calling the scene_detect filter of libavfilter to obtain the initial segmentation point, and the filter identifies the potential scene switching position by analyzing the pixel difference and histogram change between consecutive frames; to ensure algorithm consistency, the system requires the FFmpeg version to be greater than or equal to 6.1, and records ffmpeg_compat:v6.1+ in the compatibility_version field to ensure the reproducibility of the segmentation results in different environments.

[0087] (2) Through the secondary verification of the HSV histogram difference, the Bhattacharyya distance of the HSV histogram of consecutive frames is calculated, and when the distance is greater than a preset distance, such as 0.5, and the duration exceeds a preset time length, such as 0.5 seconds, the effectiveness of the segmentation boundary is confirmed, and this verification mechanism can filter out false segmentation caused by light changes or camera shaking.

[0088] (3) Dynamically merging micro-scene segments with a duration less than a preset duration, by analyzing the visual similarity and time continuity of adjacent segments, the too short segments are merged into the most relevant adjacent segments in terms of semantics, ensuring that each subblock contains sufficient semantic information.

[0089] The double-threshold scene segmentation algorithm used in the embodiments of the present application reduces the false segmentation rate by 37% compared with the single FFmpeg scheme, while maintaining a high consistency between the subblock boundaries and the semantic structure of the video.

[0090] For audio data, based on the spectral analysis results of fast Fourier transform and the silence detection information based on energy threshold, the subblocks are segmented, and when the spectral feature variation coefficient is greater than a preset spectral feature variation coefficient threshold, such as 0.4, or the silence duration is greater than a preset duration, such as 2 seconds, the subblock boundary is set.

[0091] For vector data, the K-means clustering results and dimension distribution characteristics are used to organize vectors with a cosine similarity greater than a set similarity, for example, 0.8, in the same block, and a balance constraint of the block size in the range of 64 MB-256 MB is ensured through dynamic adjustment.

[0092] For the vector data optimization strategy, a selection strategy is made according to an evaluation matrix composed of dimension size, sparsity distribution, accuracy requirement, data distribution characteristics and storage budget constraints, which specifically includes:

[0093] Based on the collaborative determination of the dimension size and the sparsity distribution characteristic parameters, the storage format type is determined.

[0094] Based on the matching analysis of the accuracy requirement and the dimension size characteristic parameters, a quantization compression or dimension reduction storage scheme is selected, and a complete set of dimension reduction reconstruction parameters is left in the column-level metadata.

[0095] Specifically, based on the collaborative determination of the dimension size and the sparsity distribution characteristic parameters, the storage format type can include: when the vector dimension is greater than a first preset dimension, for example, 1000, and the sparsity distribution characteristic parameter is less than a first preset sparsity value, for example, 0.1, a dense storage format is selected and 8bit / 16bit quantization compression is applied; when the sparsity distribution characteristic parameter is greater than a second preset sparsity value, for example, 0.7, a CSR or CSC sparse format is selected.

[0096] Based on the matching analysis of the accuracy requirement and the dimension size characteristic parameters, a quantization compression or dimension reduction storage scheme is selected, and a complete set of dimension reduction reconstruction parameters is left in the column-level metadata, which specifically includes: when the accuracy requirement allows approximate calculation and the dimension is greater than a second preset dimension, for example, 512, the dimension is reduced to a preset percentage of the original dimension, for example, 60%-70% of the original dimension, through principal component analysis, and a complete set of PCA reconstruction parameters is saved in the column-level metadata, including a kxd principal component matrix P (k is the dimension after dimension reduction, d is the original dimension), a k-dimensional eigenvalue vector λ, a d-dimensional original data mean vector μ and a reconstruction accuracy threshold ε (the default value is 0.95), to ensure that the original vector space can be reconstructed through inverse transformation x rc =P T ×x re +μ when the reconstruction error ||x ori -x re ||2is less than ε, and the accuracy control is supported. Wherein, x rc is the vector obtained by dimension reduction reconstruction, that is, the approximate original vector based on the data restoration after dimension reduction, x re is the low-dimensional vector obtained after principal component analysis (PCA) dimension reduction, x ori is the original vector, and P Tis the transpose matrix of P, ||·||2 is the L2 norm (Euclidean distance) used to quantify the difference between the original vector and the reconstructed vector.

[0097] In the embodiment of the application, the quantization strategy selection for optimizing the compression of vector data is based on the identification result of the distribution characteristics of the vector data by Kolmogorov-Smirnov test: when the test result shows that the data conforms to the normal distribution (p value is greater than a preset significance level, such as 0.05), a scalar quantization strategy is adopted, and the quantization parameters (including quantization step, zero point offset and dynamic range) are recorded in the page-level metadata; when it is identified as a multi-peak distribution, a clustering quantization strategy based on Gaussian Mixture Model (GMM) is adopted, and the clustering center matrix, covariance matrix and mixing weight parameters are saved in the column-level metadata simultaneously; when the vector is a high-dimensional dense vector (i.e. the vector dimension is greater than a preset high-dimensional threshold, such as 2048), a product quantization strategy is adopted to decompose the vector into 8 sub-vectors for quantization, and the sub-vector division boundary, each sub-space codebook and the upper limit of quantization error are recorded in the page-level metadata.

[0098] In the embodiment of the application, the native encoding reservation strategy directly reserves the native encoding structure of the compressed media data by skipping the general compression step, avoiding the problems of increased computational overhead and decreased storage efficiency caused by traditional Parquet secondary compression. The core implementation mechanism of this strategy is to identify the encoding format of the media data through the file header magic number when data is written, and record the encoding parameters, compression ratio and quality information in the metadata, so that the query engine can directly use these information for optimized access.

[0099] In the embodiment of the application, the boundary partitioning strategy based on modal characteristics realizes the transition from traditional Parquet fixed-size partitioning to semantic boundary partitioning, organizes data related in semantics in the same row group through the specific partitioning algorithms of the foregoing modal data, reduces boundary redundancy and supports more accurate selective access. The data unit generated by this strategy strictly corresponds to the row group structure of Parquet, each semantic unit is stored as an independent row group, and the data pages in the row group are organized in a physically continuous manner, ensuring the locality of the storage location of the semantic unit.

[0100] In the embodiment of the present application, the vector data optimization strategy makes full use of the mathematical characteristics of high-dimensional data to realize targeted storage and calculation optimization. The present application identifies the characteristics of the vector data through the VECTOR modal perception logic, and then applies special optimization techniques. For high-dimensional dense vectors, the system can choose to apply dimension reduction storage, which reduces the storage space while retaining the main information through principal component analysis or random projection method, and records the dimension reduction parameters in the metadata to support approximate query. It needs to be specially pointed out that the dimension reduction technique changes the semantic structure of the original vector space, so this optimization is only suitable for approximate search scenarios and is not suitable for applications that require accurate vector matching. For sparse vectors, the system uses sparse storage formats such as compressed sparse row (CSR) or compressed sparse column (CSC) to store only non-zero elements and their indexes. To solve the index reconstruction problem in block storage, the system records the global_col_offset field in the row group level metadata, which saves the column index offset of the row group in the global matrix, ensuring that the global index can be correctly reconstructed when querying across row groups. This optimization maintains the accuracy of the vector and significantly reduces the storage space requirement. For large-scale vector sets, the system supports quantization techniques, including scalar quantization and product quantization, which compress high-precision floating-point data into low-precision integer representations, significantly reducing storage requirements with acceptable accuracy loss. For vectors with specific distribution characteristics, the system also supports distribution-based encoding methods, such as statistical encoding for normally distributed vectors and dictionary encoding for sparse distributed vectors. These special optimizations enable the system to efficiently store and query vector data under different accuracy requirements and application scenarios, providing strong support for machine learning and artificial intelligence applications.

[0101] In the embodiment of the present application, the third decision mechanism serves as the final confirmation link of the storage optimization strategy, which is used to complete strategy selection, parameter calibration and execution plan formulation in combination with the expected performance evaluation results. The core logic is to ensure that the optimization strategy balances access efficiency and data accuracy while meeting storage requirements through multi-dimensional performance characterization value quantification evaluation. The specific implementation process is as follows:

[0102] (1) Construction of expected performance evaluation system

[0103] The expected performance evaluation result is determined based on the transformation effect characterization value of the data item, which includes three types of core indicators to form a multi-dimensional evaluation matrix: storage efficiency indicators, including compression ratio (compressed data volume / original data volume), metadata redundancy (metadata volume / total storage volume), used to measure the saving effect of the strategy on storage resources;

[0104] Access performance indicators: including random access delay (single data retrieval response time), batch read throughput (data volume that can be processed per unit time), reflecting the impact of the strategy on data read-write efficiency;

[0105] Precision maintenance indicators: for image, vector, etc. modalities, including reconstruction error rate (such as L2 norm error of vector reconstruction / original vector module length), feature loss degree (such as image key feature point retention ratio), to ensure that the compression or dimensionality reduction processing does not exceed the preset precision threshold.

[0106] (2) Execution flow of the third-level decision mechanism

[0107] Strategy candidate pool construction: collect candidate optimization strategies output by the first two-level decision mechanisms, such as “quantization + dimensionality reduction” combination strategies for vector data, “block encoding + adaptive compression” strategies for image data, etc., and pre-play the expected performance of each strategy based on the representation value calculation model;

[0108] Threshold comparison and adaptability judgment: compare the pre-played representation value with the system preset modality-specific threshold, such as vector data compression rate ≥ 30% and image reconstruction error rate ≤ 5%, and select strategies that meet the basic requirements;

[0109] Multi-objective optimization verification: for strategies that pass the threshold screening, use a weighted scoring method (weights are dynamically adjusted according to business scenarios, such as focusing on storage efficiency in storage-intensive scenarios and focusing on access performance in real-time retrieval scenarios) to calculate the comprehensive score, and select the optimal strategy;

[0110] Execution plan development: for the confirmed optimal strategy, specify the execution details, including:

[0111] Execution order of encoding / compression operations (such as dimensionality reduction first and then quantization to avoid precision loss accumulation);

[0112] Metadata synchronization timing (such as page-level metadata being updated in real time with data page generation, and column-level metadata being written after batch processing is completed);

[0113] Fault tolerance mechanism such as rollback plan when strategy execution fails, preferentially using the suboptimal strategy that passed the previous verification.

[0114] Through the above process, the third-level decision mechanism not only ensures the technical feasibility of the optimization strategy, but also realizes the dynamic balance of storage efficiency, access performance, and data precision, finally outputting an executable execution plan.

[0115] In the embodiment of the application, the data optimized by the modality optimization encoding strategy component is stored to the underlying physical storage layer, maintaining complete compatibility with traditional Parquet BINARY / BYTE_ARRAY columns.

[0116] Further, the modal optimization encoding strategy component generates feedback information based on the optimization result within a preset time threshold after performing targeted storage optimization and feeds back to the modal-aware format extension component, the feedback information including actual compression ratio, encoding execution time, storage efficiency measured value, and actual position of block boundary.

[0117] The modal-aware format extension component updates the metadata of the corresponding level based on the received feedback information, specifically including: when the storage efficiency measured value of the modal data in the current storage process is better than the storage efficiency benchmark value of the same modal data storage last time, dynamically adjusting the storage strategy parameters corresponding to the modal data based on the current feedback information, and synchronously updating the performance statistical information in the row group level metadata, the encoding method identifier and parameter configuration in the column level metadata, and the block boundary and semantic label in the page level metadata, to ensure that the metadata is consistent with the optimal storage state. In the embodiment of the application, the modal-aware format extension component and the modal optimization encoding strategy component establish a deep cooperative relationship through standardized bidirectional information interaction to form a complete feedback loop. Among them, the modal type information (including type_name, type_parameters, compatibility_flags three core fields) recognized by the logical type system extension component directly guides the selection decision of the native encoding preservation strategy, ensuring that the most suitable processing method is adopted for JPEG, MP3, H.264 and other compressed media data; the statistical feature information and distribution characteristic provided by the metadata structure enhancement support the boundary decision of the GOP partition strategy based on the FFmpeg scene_cut threshold, which calls the av_frame_get_side_data function defined in the FFmpeg API document to obtain the scene cut score, so that the block boundary matches the semantic structure and access mode of the data; the dimension and sparsity information of the vector data guides the vector data to select the most suitable storage format and index structure through special optimization.

[0118] The modal optimization encoding strategy component feeds back the optimization result to update the metadata structure of MAFE after executing the encoding strategy, specifically including recording performance indicators such as actual compression ratio, encoding execution time, and storage efficiency measured value in the row group level metadata, updating the encoding method identifier (such as native_encoding_preserved=true) and optimization parameter configuration in the column level metadata, and recording the actual position of the block boundary and the semantic label in the page level metadata. For vector dimension reduction storage, the system saves the key parameters of the PCA transformation matrix (principal component vector, eigenvalue, mean vector) in the column level metadata and records the target dimension after dimension reduction and the reconstruction accuracy threshold in the page level metadata, to ensure that the original vector space can be reconstructed through inverse PCA transformation for approximate matching calculation during query.

[0119] The four-layer metadata structure realizes a complete information management system from global to local through hierarchical organization. The file-level metadata provides a global modal distribution view and compatibility version information, supporting fast filtering and compatibility judgment at the file level. The row group-level and column-level metadata support medium-granularity query optimization and data skipping, providing accurate basis for predicate pushdown and column pruning. The page-level metadata realizes the finest-granularity data positioning and feature description, realizes constant-time positioning based on the row group offset table OffsetIndex, this mechanism follows the index design of Parquet format specification v2.9, detects scene switching by calling the scene_detect filter in the libavfilter library of FFmpeg, obtains the scene switching score and stores it in the page-level metadata, and supports fast frame positioning and conditional filtering in the H.264 video scene. This hierarchical design ensures that the query engine can select appropriate granularity metadata according to the query requirements, and realize multi-level optimization.

[0120] The compatibility guarantee mechanism runs through the design and implementation process of the whole technical solution. At the format level, by adding the compatibility_version field and format identification information in the file-level metadata, the system can backward compatible with traditional Parquet readers, ensuring that they can normally access the underlying data content without being affected by the added functions. At the storage level, the added modal characteristic information is stored as an optional extension in the enhanced metadata structure, without changing the original physical storage layout and data organization method. The underlying physical storage layer completely maintains the format compatibility with the traditional Parquet BINARY / BYTE_ARRAY column, ensuring the portability of data and interoperability with existing tools. At the upgrade level, the system supports a gradual upgrade path, allowing users to gradually adopt new format features according to actual needs and system capabilities, reducing migration risks and implementation complexity.

[0121] Further, the system provided by the embodiment of the present application further comprises a data reading and query optimization module, which comprises a metadata analysis unit, a query optimization unit and a data access unit.

[0122] The metadata analysis unit is used to analyze the modal type identification, feature parameters and storage strategy information layer by layer from the file-level metadata to the page-level metadata based on the four-layer hierarchical metadata architecture.

[0123] The query optimization unit is used to dynamically select the adaptive index strategy based on the modal features of the data to be queried and the query conditions, wherein when the query efficiency of the data to be queried is greater than a preset efficiency value, a high-level index strategy is adopted, otherwise a basic filtering strategy is adopted; wherein the preset efficiency value can be 0.5.

[0124] The data access unit is used for realizing efficient data access in combination with storage layer optimization characteristics, and specifically includes: realizing data zero-copy access through a native encoding reservation mechanism, supporting accurate positioning of data regions based on semantic boundary blocking, and supporting SIMD parallel computation to accelerate the query process by using a vector optimized storage format.

[0125] In the embodiment of the application, the storage efficiency is normalized storage efficiency, and the query efficiency is normalized query efficiency.

[0126] Further, the system provided by the embodiment of the application can realize data writing processing flow and data reading and query optimization flow in an actual application scenario.

[0127] The data writing flow embodies the cooperative working mechanism of the two core technical components, and realizes a complete conversion process from original multi-modal data to optimized storage. The flow is divided into two main stages of semantic analysis and storage optimization, ensuring that each data item can obtain the most suitable processing mode.

[0128] The timing control of the interaction of the two components adopts a synchronous coordination mechanism, ensuring the integrity of the semantic analysis result and the accuracy of the encoding strategy selection. In the initial stage of data writing, the system first performs hardware environment verification, calculates the matching degree of the current hardware fingerprint with the standard test environment, and triggers the reference value recalibration flow when the matching degree is lower than 90%, ensuring that the C i baseline The effectiveness of the reference value. The modal perception format expansion component then performs batch preprocessing on the input data, establishing a preliminary classification and feature extraction queue of the data items. In the processing process, the system calculates the transformation effect representation value of each data item in real time. When the semantic analysis of a single data item is completed, the system immediately delivers the standardized output result to the modal optimization encoding strategy component. This instant delivery mechanism avoids the need for caching a large number of intermediate results, while ensuring that the encoding strategy can make decisions based on the latest analysis results.

[0129] The interaction of the two components follows a strict bidirectional timing protocol, ensuring the consistency, reliability and complete feedback loop of data processing. In the semantic analysis stage, the modal perception format expansion component generates a standardized semantic descriptor after completing the semantic analysis of a single data item. In the strategy decision stage, the modal optimization encoding strategy component must return an encoding execution plan within the dynamically calculated response timeout threshold after receiving the semantic descriptor, and the threshold is determined by a hardware adaptive adjustment mechanism, with a minimum value of 150 milliseconds. The encoding execution plan includes specific encoding methods, parameter configurations and expected performance indicators.

[0130] In the feedback update phase, the modal optimization encoding strategy component must feed back the optimization results to the modal perception format expansion component for metadata update within 50 milliseconds after the completion of encoding strategy execution. The feedback information includes key indicators such as actual compression ratio, encoding execution time, storage efficiency measured value, and block boundary actual position. The modal perception format expansion component updates the corresponding level metadata immediately after receiving the feedback, including: row group level metadata update performance indicators and statistical information, column level metadata update encoding method identifier and parameter configuration, page level metadata update block boundary and semantic label. For vector dimension reduction storage, the modal perception format expansion component saves the complete PCA reconstruction parameters in the column level metadata to ensure the vector space reconstruction capability during query.

[0131] The system sets up a complete exception handling and synchronization mechanism. If the modal optimization encoding strategy component response times out, the system automatically enables the default encoding strategy and records the timeout event in the compatibility version field. If the feedback update fails, the system marks the data item as "metadata inconsistency" state and triggers the reprocessing flow. After processing 100 data items, the two components perform a bidirectional state synchronization check to ensure the consistency of the forward processing and feedback update progress.

[0132] In the semantic analysis phase, the modal perception format expansion component performs deep analysis on the input multi-modal data. The logical type system extension determines the specific modal type through data feature recognition, such as identifying JPEG image format through file header magic number, identifying audio format through sampling rate and channel information, and identifying vector data through array dimension and data type, and extracts corresponding format parameters, resolution parameters, encoding parameters, and other key features. The metadata structure enhancement records the recognition results in a four-layer architecture. File level metadata records global modal distribution and version compatibility information, row group level metadata records modal distribution characteristics and correlation mapping within the row group, column level metadata records modal specific attributes and compression strategy information, and page level metadata records precise positioning and optimization information of the data within the page.

[0133] In the storage optimization stage, the modal optimization encoding strategy component performs targeted optimization processing according to the semantic analysis results. The native encoding retention strategy checks the existing encoding format of the media data. For compressed formats such as JPEG, PNG, MP3, AAC, H.264, H.265, etc., the internal encoding structure is directly retained to avoid quality loss and storage efficiency decline caused by repeated compression. The boundary block strategy based on modal characteristics uses corresponding block methods according to the semantic characteristics of different modalities. The pyramid block strategy is used for image data to support multi-resolution access, the scene block based on GOP structure is used for video data to maintain encoding integrity, the block based on spectral characteristics is used for audio data to correspond to auditory perception, and the clustering block based on similarity is used for vector data to improve query efficiency. The vector data is specially optimized to select the most suitable storage format according to the dimension, sparsity and distribution characteristics of the vector, such as quantization compression for dense vectors and CSR / CSC format for sparse vectors. Finally, all optimized data is written to the underlying physical storage layer, strictly maintaining compatibility with the traditional ParquetBINARY / BYTE_ARRAY column format.

[0134] Taking H.264 encoded video as an example, the input data is a standard video file with a code rate of 8Mbps. The decision-making process of the modal optimization encoding strategy (MOES) first identifies the codec type as h264, and then detects the integrity of the NAL_unit structure. When the structure is complete, the encoding_flags flag is set to 0x8F, indicating that the system skips the Snappy compression layer and directly retains the native encoding structure. The actual effect of this embodiment shows that, compared with the traditional Parquet secondary compression method for the same video data, the native encoding retention strategy reduces the write delay by 62%, and avoids the video quality loss caused by repeated compression. This embodiment verifies the effectiveness and practicality of the native encoding retention strategy in practical application.

[0135] The data reading process realizes an efficient query processing mechanism based on modal perception technology, and improves the query performance and access efficiency of multi-modal data through hierarchical metadata analysis and modal-specific optimization strategies. The process includes three key stages of metadata analysis, query optimization and data access. The specific implementation mechanisms include: the metadata analysis stage analyzes modal characteristic information from file level to page level through a four-layer metadata architecture, the query optimization stage dynamically selects high-level index strategy or basic filtering strategy based on query efficiency greater than the preset efficiency value, and the data access stage realizes zero-copy access through native encoding retention, supports precise positioning through semantic boundary block, and supports SIMD (Single Instruction Multiple Data) parallel computing through vector optimization format.

[0136] In the metadata parsing phase, the system obtains the semantic information required by the query in four layers of architecture. File-level metadata parsing quickly obtains global modal distribution information and version compatibility, helping the query engine to determine the modal type and data range involved in the query, and realizing coarse-grained filtering at the file level. Row group-level metadata parsing determines the relevant data blocks and modal distribution characteristics, supporting data skipping and selective reading based on modal correlation. Column-level metadata parsing extracts modal-specific parameters and encoding strategy information, providing accurate basis for subsequent predicate pushdown and column pruning. Page-level metadata parsing obtains precise positioning information and feature summaries of the data within the page, supporting the most fine-grained selective access and conditional filtering.

[0137] In the query optimization phase, the system uses modal perception technology to perform multi-level query optimization strategies, with normalized query efficiency playing a key role in dynamic cost estimation. The dynamic cost estimation mechanism is realized through normalized query efficiency in the query optimization phase. When the normalized query efficiency of a certain modal type is greater than 0.5, high-level index strategies are enabled, including the combined use of multi-dimensional index, vector index and composite index, otherwise basic filtering strategies are used, only using a single dimension of simple index. This threshold is determined based on ROC curve analysis of 100,000 queries, when the normalized query efficiency is greater than 0.5, the positive probability of advanced index is greater than 95%. The system also considers the historical trend of the transformation effect value of each data item, when the average transformation effect value of a certain modal data continues to rise, the system will dynamically adjust the business weight coefficient of this modal to configure the query optimization strategy weight, realizing adaptive performance optimization. This dynamic selection mechanism is based on historical query performance data and real-time system load conditions, ensuring accurate matching of query optimization strategies and actual performance benefits.

[0138] Intelligent predicate pushdown based on modal-specific statistical information, such as image resolution, color space parameters for image filtering, audio sampling rate, bit depth for audio filtering, vector dimension, sparsity for vector query optimization. According to the semantic boundary information established by the intelligent block strategy, the query engine can accurately locate the data pages and row groups that need to be accessed, avoiding unnecessary I / O operations and data transmission.

[0139] For vector similarity queries, the system implements a complete multi-level query support mechanism. For dimension-reduced stored vector data, the query engine first reads the PCA reconstruction parameters from the column-level metadata, and then selects the processing strategy according to the query type: for exact match queries, the original vector space is completely reconstructed through inverse PCA transformation for calculation; for approximate search queries, similarity calculation is directly performed in the reduced dimension space, and the result quality is controlled according to the reconstruction accuracy threshold; for range queries, double filtering is performed by combining distance calculation in the reduced dimension space and error estimation in the original space. For quantized stored vector data, the system performs corresponding dequantization operations according to the quantization type: scalar quantization restores floating-point values through inverse PCA transformation; clustering quantization reconstructs the probability distribution by looking up the cluster center and covariance matrix; product quantization reconstructs the complete vector by combining the codebooks of each subspace.

[0140] In the data access phase, various optimization strategies work together to improve access efficiency, which is reflected in three levels of collaborative optimization. The native encoding reservation strategy enables the system to directly access compressed media data without additional decompression and re-encoding overhead, especially for JPEG images and H.264 videos, which can achieve zero-copy access. The semantic boundary information of the boundary block strategy based on modal features supports accurate data positioning and selective reading, ensuring that only query-related data blocks are accessed. The storage format optimized for vector data supports efficient vector operations and batch processing, accelerating vector operations through SIMD instructions and parallel computing. The entire query processing flow from metadata parsing to data return forms a complete optimization chain, achieving performance improvement in multi-modal data query.

[0141] The actual effect of the technical scheme is further verified based on the experimental data of the self-built multi-modal test data set (note: the following data is example data for verifying the technical scheme). For 4K video data, the storage size of the traditional Parquet scheme is 450MB, and the query delay is 1200 milliseconds, while the storage size of the present scheme is 318MB, and the query delay is 380 milliseconds, the compression ratio is improved by 1.42 times, the query performance improvement coefficient γ=1200 / 380≈3.16 times, the normalized storage efficiency σ=max(0,1-318 / 450)≈0.29, the normalized query efficiency T=(3.16-1) / (3.16+1)×H(3.16-1)≈0.52. In the query-intensive scenario (α=0.15, β=0.8, δ=0.05), the transformation effect representation value G=0.15×0.29+0.8×0.52+0.05×0.29×0.52≈0.46.

[0142] For medical image DICOM data, the storage size of the traditional scheme is 3.2 GB, and the query delay is 860 milliseconds. The storage size of the present scheme is 1.4 GB, and the query delay is 480 milliseconds. The compression ratio is improved by 2.29 times, the query performance improvement coefficient γ = 860 / 480 ≈ 1.79 times, the normalized storage efficiency σ = max(0, 1-1.4 / 3.2) ≈ 0.56, and the normalized query efficiency T = (1.79-1) / (1.79+1) × H(1.79-1) ≈ 0.28. In the storage sensitive scene (α = 0.75, β = 0.2, δ = 0.05), the reform effect representation value G = 0.75 × 0.56 + 0.2 × 0.28 + 0.05 × 0.56 × 0.28 ≈ 0.49. These experimental data show that the technical scheme can realize significant performance improvement in different types of multi-modal data processing and different business scenarios, and verifies the practicability and effectiveness of the improved mathematical model and technical scheme.

[0143] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, the steps described in the present application can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical scheme disclosed in the present application can be achieved, which is not limited herein.

[0144] The above specific embodiments do not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A multi-modal data augmentation storage system based on Parquet format, characterized in that, The system realizes unified storage and optimization of multi-modal data through native modification and extension of the Parquet format, including a modal-aware format extension component and a modal optimization encoding strategy component; The modal-aware format extension component realizes deep integration of multi-modal data semantics and Parquet storage structure through targeted extension of the logical type system and metadata architecture of the Parquet format, thereby establishing a semantic description system for multi-modal data, including a logical type system extension submodule and a metadata structure enhancement submodule; The logical type system extension submodule realizes extension of the logical type definition of the Parquet format by inheriting the original LogicalType base class of the Parquet format and adding modal-specific parameter fields; The metadata structure enhancement submodule is configured to construct a four-layer metadata architecture at the file level, row group level, column level, and page level, which is accurately mapped to the page structure of the Parquet physical storage and is used to record the global distribution, local features, and fine positioning information of multi-modal data at different levels, thereby realizing deep coupling of modal feature information and the Parquet storage structure; The modal optimization encoding strategy component realizes instruction-level fusion with the Parquet compression algorithm and is configured to receive the semantic descriptor generated by the modal-aware format extension component, perform targeted storage optimization based on the type characteristics, feature parameters, and optimization suggestions of multi-modal data.

2. The system of claim 1, wherein, The logical type system extension submodule specifically includes: A modal type identification unit configured to identify the modal types of each data item in multi-modal data and map them to the extended multi-modal logical type; A feature parameter extraction unit configured to extract the feature parameters of the corresponding modal based on the identified modal type and associate the extracted feature parameters with the semantic description field of the logical type; An optimization suggestion generation unit configured to generate storage strategy optimization suggestions based on the feature parameters and form a standardized semantic descriptor containing the modal type, feature parameters, and optimization suggestions.

3. The system of claim 1, wherein, The four-layer metadata architecture includes a file-level metadata storage unit, a row group-level metadata storage unit, a column-level metadata storage unit, and a page-level metadata storage unit. The file-level metadata storage unit is configured to extend the FileMetaData structure of the Parquet format to provide a global view and cross-modal statistical information. The row group-level metadata is configured to extend the RowGroup structure of the Parquet format to describe data block organization and modal distribution characteristics. The column-level metadata is configured to extend the ColumnChunk.meta_data field of the Parquet format to provide specific information at the column dimension and modal characteristic description. The page-level metadata realizes hybrid storage using the extension area of the PageHeader of the Parquet format and the reserved area at the end of the data page to store the positioning information of the data within the page.

4. The system of claim 3, wherein, The page-level metadata storage rule includes: fixed-length fields are stored in the PageHeader extension area, and a fixed offset design is adopted to support O(1) time complexity field positioning; variable-length fields are stored in the reserved area at the tail of the data page, and Length-Prefixed coding is adopted, the first 4 bytes store the data length, and the subsequent bytes store the actual data content, wherein each metadata page contains a 32-byte fixed header, which stores the offset table of the key field, and the field value is stored in little-endian sequence, which is consistent with the Parquet format specification v2.

9.

5. The system of claim 1, wherein, The modal optimization encoding strategy component includes: A native encoding reservation submodule is configured to reserve native encoding for compressed media data to avoid secondary compression; A modal feature-based boundary block submodule is configured to implement semantic boundary block corresponding to a Parquet row group structure based on modal features; A vector data optimization submodule is configured to use a sparse format or quantization compression adapted to Parquet columnar storage for vector data.

6. The system of claim 1, wherein, The modal optimization encoding strategy component processes the semantic descriptor through a three-level decision mechanism to achieve accurate selection and configuration of the encoding strategy, wherein: The first-level decision mechanism is configured to determine the applicable encoding strategy category based on the modal type identifier and the data integrity check result; The second-level decision mechanism is configured to finely adjust the strategy parameters based on specific values in the modal feature parameter set; The third-level decision mechanism is configured to make a final strategy confirmation and develop an execution plan in combination with the expected performance evaluation result.

7. The system of claim 6, wherein, The second-level decision mechanism specifically includes: For the native encoding reservation strategy, the existing compression state indicator in the modal feature parameter is used for fine judgment, and when the compression ratio exceeds the preset threshold and the quality loss evaluation is within an acceptable range, it is determined that the data item is suitable for native encoding reservation; For the modal feature-based boundary block strategy, the structured information in the modal feature parameter and the access prediction model based on historical query patterns are used to determine the block boundary; For the vector data optimization strategy, the evaluation matrix composed of dimension size, sparsity distribution, precision requirement, data distribution feature, and storage budget constraint is used for strategy selection.

8. The system of claim 7, wherein, For the native encoding reservation strategy, the existing compression state indicator in the modal feature parameter is used for fine judgment, and when the compression ratio exceeds the preset threshold and the quality loss evaluation is within an acceptable range, it is determined that the data item is suitable for native encoding reservation, specifically including: For image data, whether to reserve its native encoding structure is determined based on its quality-related parameters and file size and original data proportion feature parameters; For video data, whether to reserve its native encoding format is determined based on its code rate control mode and encoding structure integrity feature parameters; For audio data, whether to reserve its native frame structure is determined based on its bit rate parameter and encoding standard compliance feature; For the modal feature-based boundary block strategy, the structured information in the modal feature parameter and the access prediction model based on historical query patterns are used to determine the block boundary, specifically including: For image data, based on its size characteristic parameters and pixel density variation characteristic parameters, a corresponding block strategy is selected; For video data, a double-threshold scene segmentation strategy is adopted to realize semantic boundary recognition, which includes: based on scene switching detection for preliminary segmentation, combined with inter-frame feature difference degree for secondary verification to confirm boundary effectiveness, dynamic merging is performed on too short segments to ensure the semantic integrity of the block; For audio data, based on the spectral feature variation parameter and the silent duration feature parameter, the block boundary is set; For vector data, based on the clustering results and the dimension distribution characteristics, the block is organized by combining the vector similarity characteristic parameters, and the balance of the block size is ensured by dynamic adjustment; For vector data optimization strategy, according to the evaluation matrix composed of dimension scale, sparsity distribution, accuracy requirement, data distribution characteristics and storage budget constraints, the strategy selection is carried out, which specifically includes: Based on the collaborative judgment of dimension scale and sparsity distribution characteristic parameters, the storage format type is determined; Based on the matching analysis of accuracy requirement and dimension scale characteristic parameters, the quantization compression or dimension reduction storage scheme is selected, and the complete dimension reduction reconstruction parameter set is reserved in the column-level metadata.

9. The system of claim 7, wherein, After the modal optimization encoding strategy component performs targeted storage optimization, it generates feedback information based on the optimization results within a preset time threshold and feeds it back to the modal awareness format expansion component, the feedback information includes actual compression ratio, encoding execution time, storage efficiency measured value, block boundary actual position; The modal awareness format expansion component updates the metadata of the corresponding level based on the received feedback information, specifically including: when the storage efficiency measured value of the modal data in the current storage process is better than the storage efficiency benchmark value of the same modal data storage last time, dynamically adjust the storage strategy parameters of the modal data based on the current feedback information, and synchronously update the performance statistical information in the row group level metadata, the encoding method identifier and parameter configuration in the column level metadata, and the block boundary and semantic label in the page level metadata, to ensure that the metadata is consistent with the optimal storage state.

10. The system of claim 7, wherein, It also includes a data reading and query optimization module, which includes a metadata analysis unit, a query optimization unit and a data access unit; The metadata analysis unit is used to analyze the modal type identification, characteristic parameters and storage strategy information layer by layer from file-level metadata to page-level metadata based on the four-layer hierarchical metadata architecture; The query optimization unit is used to dynamically select the adaptive index strategy based on the modal characteristics of the data to be queried and the query conditions, wherein when the query efficiency of the data to be queried is greater than the preset efficiency value, the advanced index strategy is adopted, otherwise the basic filtering strategy is adopted; The data access unit is used to realize efficient data access by combining the storage layer optimization characteristics, specifically including: realizing data zero-copy access through native encoding reservation mechanism, supporting accurate positioning of data area based on semantic boundary block, and supporting SIMD parallel computing to accelerate query process by using vector optimization storage format.

Images