Feature quantification method and apparatus, and non-volatile storage medium and electronic device

Abstract

Disclosed in the present invention are a feature quantification method and apparatus, and a non-volatile storage medium and an electronic device. The method comprises: acquiring a first feature of a first type, wherein the first feature comprises first feature data in a plurality of dimensions; quantifying the first feature by using a preset quantification dictionary, so as to obtain a reconstructed feature, wherein the preset quantification dictionary comprises a preset number of sample features, each sample feature comprising consecutive feature codes, and sample feature data in a plurality of dimensions, the reconstructed feature comprises reconstructed feature data in a plurality of dimensions, reconstructed feature data in a target dimension is a feature code of a target sample feature, and the target sample feature is a sample feature that is closest to first feature data and sample feature data in the target dimension; and determining a second feature of a second type according to the reconstructed feature. By means of the present invention, the technical problem of feature data occupying a relatively large number of hardware resources is solved.

Description

Feature quantization method, device, non-volatile storage medium and electronic device

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application is based on the application with CN application number 202311112293.8 and application date August 30, 2023, and claims its priority. The disclosed content of the CN application is hereby introduced as a whole into this application. Technical Field

[0003] The present invention relates to the field of artificial intelligence, and in particular to a feature quantization method, device, non-volatile storage medium, and electronic device. Background Art

[0004] With the rapid development of AI (artificial intelligence) chips, edge AI cameras are now widely used in various fields. However, AI cameras still lag significantly behind servers and personal computers in terms of hardware resources, with limited computing power and storage. However, AI cameras offer advantages such as low cost and easy deployment, making them a promising market.

[0005] Current feature query and matching algorithms rely on hardware platforms like servers or PCs, which offer high computing power and massive storage capacity. These hardware devices can directly use Float (floating-point) feature data for query matching and database storage. Feature matching on AI camera platforms is in high demand, but directly migrating server-side feature matching query algorithms to AI cameras with lower hardware resources fails to meet practical requirements across various performance levels. AI cameras lack sufficient storage capacity to store large Float-type quantitative feature databases, and their low computing power significantly reduces response sensitivity when performing Float feature query matching.

[0006] Currently, no effective solution has been proposed to address the problem that the above-mentioned feature data occupies a large amount of hardware resources.

[0007] Summary of the Invention

[0008] According to one aspect of an embodiment of the present invention, a feature quantization method is provided, comprising: obtaining a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one to the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of a target dimension is a feature code of a target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; determining a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0009] In some embodiments, before using a preset quantization dictionary to quantize the first feature to obtain a reconstructed feature, the method further includes: obtaining multiple training features of the first type, wherein the training features include training feature data of multiple dimensions; clustering the multiple training features to obtain the preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; and assigning continuous feature codes to the preset number of sample features to generate the preset quantization dictionary.

[0010] In some embodiments, the first feature is quantized using a preset quantization dictionary to obtain a reconstructed feature, including: calculating the absolute value of the difference between the first target data and each sample target data to obtain a target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; determining the target sample feature based on the sample target data with the smallest target data difference; and encoding the feature of the target sample feature to determine it as the reconstructed feature data of the reconstructed feature in the target dimension.

[0011] In some embodiments, determining the second feature of the second type based on the reconstructed feature includes: decentralizing the reconstructed feature to obtain the second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is determined according to the numerical range required by the second type.

[0012] In some embodiments, after determining a second feature of the second type based on the reconstructed feature, the method further includes: when the first feature is an original feature, calculating a feature modulus of the second feature, wherein the original feature is used to indicate construction of a quantitative feature database; and constructing the quantitative feature database based on the second feature and the feature modulus of the second feature.

[0013] In some embodiments, after determining a second feature of the second type based on the reconstructed feature, the method further includes: when the first feature is a query feature, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database, wherein the query feature is used to query the target feature in the quantitative feature database; and determining that the preset feature with the largest feature similarity is the target feature.

[0014] In some embodiments, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database includes: obtaining multiple preset features in the quantitative feature database and the feature modulus of each of the preset features; calculating the matrix multiplication between the second feature and the quantitative feature database to obtain a product result; dividing the product result by the feature modulus of each of the preset features to determine the feature similarity between the second feature and each of the preset features.

[0015] According to another aspect of an embodiment of the present invention, a feature quantization device is also provided, including: an acquisition module for acquiring a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; a quantization module for quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one to the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a determination module for determining a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0016] According to another aspect of an embodiment of the present invention, a non-volatile storage medium is provided. The non-volatile storage medium is used to store a program. When the program is running, the device where the non-volatile storage medium is located is controlled to execute the above-mentioned feature quantization method.

[0017] According to another aspect of an embodiment of the present invention, an electronic device is provided, including: a memory and a processor, wherein the processor is configured to run a program stored in the processor, wherein the program executes the above-mentioned feature quantization method when running. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0019] FIG1 is a flow chart of a feature quantization method according to an embodiment of the present invention;

[0020] FIG2 is a schematic diagram of a solution for constructing a quantitative feature database and performing feature query based on the quantitative feature database according to an embodiment of the present invention;

[0021] FIG3 is a schematic diagram of a training preset quantization dictionary according to an embodiment of the present invention;

[0022] FIG4 is a schematic diagram of feature quantization coding according to an embodiment of the present invention;

[0023] FIG5 is a schematic diagram of constructing a quantitative feature database according to an embodiment of the present invention;

[0024] FIG6 is a schematic diagram of a feature query process according to an embodiment of the present invention;

[0025] FIG7 is a schematic diagram of a feature quantization device according to an embodiment of the present invention;

[0026] FIG8 is a structural block diagram of a computer terminal according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] Embodiments of the present invention provide a feature quantization method, device, non-volatile storage medium, and electronic device to at least solve the technical problem that feature data occupies a large amount of hardware resources.

[0030] In an embodiment of the present invention, a first feature of a first type is obtained, wherein the first feature includes first feature data of multiple dimensions; the first feature is quantized using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a second feature of a second type is determined based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, and the second feature The data is determined based on the reconstructed feature data. The first type and the second type are different. The second feature data meets the numerical range required by the second type, thereby realizing the conversion of the first feature of the first type into the second feature of the second type. Since the second feature data in the second feature is determined based on the feature coding of a preset quantization dictionary, a second feature based on the coding form can be obtained. The volume of the second feature data in the coding form is smaller than that of the first feature in the data form. Therefore, in the feature processing process, the hardware resources occupied by the second feature in the coding form will be smaller than the hardware resources occupied by the first feature in the data form, thereby achieving the purpose of reducing the data volume of the feature data, realizing the technical effect of reducing the occupation of hardware resources by the feature data, and thus solving the technical problem of the large occupation of hardware resources by the feature data.

[0031] According to an embodiment of the present invention, an embodiment of a feature quantization method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0032] FIG1 is a flow chart of a feature quantization method according to an embodiment of the present invention. As shown in FIG1 , the method includes the following steps:

[0033] Step S102: Acquire a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions;

[0034] Step S104: quantize the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension;

[0035] Step S106: determining a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0036] In an embodiment of the present invention, a first feature of a first type is obtained, wherein the first feature includes first feature data of multiple dimensions; the first feature is quantized using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a second feature of a second type is determined based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, and the second feature The data is determined based on the reconstructed feature data. The first type and the second type are different. The second feature data meets the numerical range required by the second type, thereby realizing the conversion of the first feature of the first type into the second feature of the second type. Since the second feature data in the second feature is determined based on the feature coding of a preset quantization dictionary, a second feature based on the coding form can be obtained. The volume of the second feature data in the coding form is smaller than that of the first feature in the data form. Therefore, in the feature processing process, the hardware resources occupied by the second feature in the coding form will be smaller than the hardware resources occupied by the first feature in the data form, thereby achieving the purpose of reducing the data volume of the feature data, realizing the technical effect of reducing the occupation of hardware resources by the feature data, and thus solving the technical problem of the large occupation of hardware resources by the feature data.

[0037] In the above step S102 , the first type may be a floating point type, such as a Float type.

[0038] In the above step S104, the preset quantization dictionary can record multiple sample features of the first type, and the preset number of sample features recorded in the preset quantization dictionary can be determined according to the numerical range required by the second type. For example, when the second type is Int8 type, the numerical range of Int8 is [-128,127], then Unsigned Int8[0,255] can be used as the preset number of sample features stored in the preset quantization dictionary.

[0039] In some embodiments, Int8 represents an eight-bit binary code, and the maximum binary value range of Int8 is [00000000, 11111111], and the maximum decimal value range is [0, 255].

[0040] In the above step S104, the multiple sample features in the preset quantization dictionary are arranged in order, and a continuous feature code is assigned to each sample feature in turn according to the arrangement order, so that each sample feature has a unique feature code.

[0041] In the above step S104, the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension, including: determining the first feature data of the target dimension in the first feature; determining the sample feature data of each sample feature in the target dimension among the multiple sample features in the preset quantization dictionary; then comparing the first feature data with the multiple sample feature data one by one, determining the sample feature data closest to the first feature data among the multiple sample feature data, and replacing the first feature data with the feature code of the sample feature where the closest sample feature data is located, to obtain the reconstructed feature data of the reconstructed feature in the target dimension.

[0042] In some embodiments, the target dimension may be any dimension, and the first feature, the sample feature, the reconstructed feature, and the second feature represent the corresponding same dimension.

[0043] In the above step S106, the second type can be an integer, and the second feature of the second type includes second feature data of multiple dimensions, and the second feature data meets the numerical range required by the integer. For example, when the second type is Int8 type, the numerical range required by the second type is [-128,127].

[0044] In some embodiments, before using a preset quantization dictionary to quantize the first feature to obtain a reconstructed feature, the method also includes: obtaining multiple training features of the first type, wherein the training features include training feature data of multiple dimensions; clustering the multiple training features to obtain a preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; assigning continuous feature codes to the preset number of sample features to generate a preset quantization dictionary.

[0045] In the above-mentioned embodiment of the present invention, when training a preset quantization dictionary, clustering can be performed based on multiple first-type training features to obtain a preset number of clusters, and corresponding sample features are determined according to each cluster, and then continuous feature codes are assigned to the multiple sample features, thereby realizing the generation of the preset quantization dictionary.

[0046] In some embodiments, clustering multiple training features to obtain a preset number of sample features includes: clustering multiple training features to obtain a preset number of cluster clusters, wherein each cluster cluster has a corresponding cluster center point; and determining the sample features based on the cluster center point.

[0047] In some embodiments, the sample feature may be a training feature serving as a cluster center point in a cluster based on the training feature.

[0048] In some embodiments, the first feature is quantized using a preset quantization dictionary to obtain the reconstructed feature, including: calculating the absolute value of the difference between the first target data and each sample target data to obtain the target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; determining the target sample feature based on the sample target data with the smallest target data difference; encoding the feature of the target sample feature to determine it as the reconstructed feature data of the reconstructed feature in the target dimension.

[0049] In the above embodiment of the present invention, the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension. Therefore, the first target data and the sample target data are in the same dimension of the first feature and the sample feature. Then, when determining the sample target data closest to the first target data, the absolute value of the difference between the first target data and each sample target data can be calculated to obtain the target data difference, and the sample target data with the smallest target data difference can be determined. The sample feature where the sample target data is located is determined as the target sample feature, and then the feature code of the target sample feature is determined as the reconstructed feature data of the reconstructed feature in the target dimension, thereby realizing the determination of the reconstructed feature.

[0050] In some embodiments, determining the second feature of the second type based on the reconstructed feature includes: decentralizing the reconstructed feature to obtain the second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is determined according to the numerical range required by the second type.

[0051] In the above-mentioned embodiment of the present invention, the reconstructed feature data of multiple dimensions in the reconstructed feature are feature codes of the sample feature. In order to make the second feature data determined based on the reconstructed feature data meet the numerical range required by the second type, it is necessary to perform data processing on the reconstructed feature data. For example, by performing centering processing on the reconstructed feature, the second feature data generated based on the reconstructed feature data can meet the numerical range required by the second type.

[0052] In some embodiments, when the second type of the second feature is Int8, the numerical range required by the second type is [-128, 127], and the feature code of the reconstructed feature data record in the reconstructed feature belongs to Unsigned Int8 [0, 255]. In order to generate the second feature data based on the numerical range of [-128, 127] based on the feature code based on the numerical range of [0, 255], the preset constant can be set to 128, and the reconstructed feature can be centered according to the preset constant, so that the generated second feature data can meet the numerical range required by the second type.

[0053] In some embodiments, after determining a second feature of the second type based on the reconstructed feature, the method further includes: when the first feature is an original feature, calculating a feature modulus of the second feature, wherein the original feature is used to indicate the construction of a quantitative feature database; and constructing a quantitative feature database based on the second feature and the feature modulus of the second feature.

[0054] In some embodiments, the above-mentioned feature quantification method can be used to establish a quantitative feature database.

[0055] In the above embodiment of the present invention, when the first feature is the original feature for which a quantitative feature database needs to be constructed, the first feature can be quantized using the above feature quantization method to reduce the data volume of the first feature to obtain the second feature, and the feature modulus of the second feature is calculated. Then, the second feature and the corresponding feature modulus are stored in the quantitative feature database, thereby realizing the construction of the quantitative feature database, and then feature query can be performed based on the constructed quantitative feature database.

[0056] In some embodiments, the second feature may be used as a preset feature, the characteristic mode of the second feature may be used as the characteristic mode of the preset feature, and the preset feature and the corresponding characteristic mode may be stored in a quantitative feature database.

[0057] In some embodiments, after determining the second feature of the second type based on the reconstructed feature, the method also includes: when the first feature is a query feature, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database, wherein the query feature is used to query the target feature in the quantitative feature database; determining the preset feature with the largest feature similarity as the target feature.

[0058] In some embodiments, the above-mentioned feature quantification method can be used to perform feature query on a quantitative feature database.

[0059] In the above embodiment of the present invention, when the first feature is a query feature that requires a feature query on a quantitative feature database, the first feature can be quantified using the above feature quantization method to reduce the data volume of the first feature to obtain a second feature, and then the second feature is matched with a preset feature in the quantitative feature database to determine the feature similarity of each preset feature, and then the preset feature with the largest feature similarity is selected from the quantitative feature database as the target feature, thereby achieving the technical effect of querying the target feature in the quantitative feature database based on the query feature.

[0060] In some embodiments, the quantitative feature database may record multiple preset features and feature identifiers of each preset feature, such as preset ID information of the preset feature, and then, when determining the target feature, the target ID information of the target feature may be determined.

[0061] In some embodiments, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database includes: obtaining multiple preset features and the feature modulus of each preset feature in the quantitative database (i.e., the quantitative feature database); calculating the matrix multiplication of the second feature and the quantitative database to obtain the product result; dividing the product result by the feature modulus of each preset feature to determine the feature similarity between the second feature and each preset feature.

[0062] In the above embodiment of the present invention, matrix multiplication is performed on the second feature and all preset features in the quantization database, and the product result is divided by the feature modulus of each preset feature, so as to determine the feature similarity between the second feature and each preset feature, thereby realizing the determination of feature similarity.

[0063] The present invention also provides an optional embodiment, which provides a feature quantization and query matching method adapted to an AI camera platform. Using this method, query matching can be completed quickly. At the same time, storing the quantized feature database only requires less storage space, and the feature matching algorithm can be adapted to an AI camera platform with lower hardware configuration.

[0064] Figure 2 is a schematic diagram of a scheme for constructing a quantitative feature database and performing feature query based on the quantitative feature database according to an embodiment of the present invention. As shown in Figure 2, the scheme includes: a feature acquisition module, a feature quantization module, a quantitative feature database construction module and a target feature matching module.

[0065] In some embodiments, the feature acquisition module is mainly used to obtain the original features of the Float type (i.e., the first feature) on the AI ​​camera, and obtain the original features for subsequent feature quantification (such as determining the second feature), database construction (such as establishing a quantitative feature database) and target matching (such as querying target features in the quantitative feature database).

[0066] In some embodiments, the feature quantization module is used to train the original features (such as sample features) obtained on the AI ​​camera, obtain a preset quantization dictionary, quantize and decentralize the first feature using the preset quantization dictionary, and obtain a quantized feature of the Int8 type (such as the second feature).

[0067] In some embodiments, the quantitative feature database construction module is used to store the quantized Int8 type features (such as the second feature or the preset feature), build a quantitative feature database, and store it in the AI ​​camera storage.

[0068] In the above embodiment of the present invention, the memory for storing the Int8 type feature (i.e., the second type of feature, such as the second feature or the preset feature) is reduced by 4 times compared to directly storing the Float type feature (i.e., the first type of feature, such as the first feature or the original feature).

[0069] In some embodiments, the target feature matching module is used to perform Int8 quantization on the query feature (ie, the first feature), and then use the quantized feature (ie, the second feature) to directly match the quantized feature database to obtain feature information of the target feature.

[0070] In the above embodiment of the present invention, compared with matching using only original features of the Float type, the quantized features (ie, the second features) have a small amount of calculation and a more sensitive matching response.

[0071] In some embodiments, the features of the Float type are quantized and encoded using Int8 (i.e., the first feature of the first type is converted into the second feature of the second type), and a quantized feature database is established. The specific processing steps are as follows:

[0072] S21. Use the Float type feature to train the preset quantization dictionary.

[0073] Figure 3 is a schematic diagram of a training preset quantization dictionary according to an embodiment of the present invention. As shown in Figure 3, it is assumed that the training data of the preset quantization dictionary has M Float features (that is, M training features), each training feature has L training feature data, and clustering is performed in the M dimension to form 256 center points (equivalent to obtaining 256 sample features), and finally a preset quantization dictionary of size 256×L is obtained.

[0074] S22: quantize the first feature using a preset quantization dictionary.

[0075] Figure 4 is a schematic diagram of a feature quantization coding according to an embodiment of the present invention. As shown in Figure 4, the first feature of the L dimension is subtracted from the preset quantization dictionary and the absolute value is taken. The row number of the minimum value in each column is extracted (equivalent to determining the feature coding of the target sample feature) to determine the reconstructed feature. There are L-dimensional data in total, and the numerical range of the feature coding is between [0, 255].

[0076] S23. Reconstruct the decentralized features to obtain Int8 features.

[0077] In some embodiments, as shown in FIG4 , the reconstructed features are decentered after subtracting a preset constant 128, and the data is quantized into Int8 data with a value range of [-128, 127].

[0078] S24. Build a quantitative feature database.

[0079] FIG5 is a schematic diagram of a method for constructing a quantitative feature database according to an embodiment of the present invention. As shown in FIG5 , the N original features (i.e., the first features) of the database are quantized using the above-mentioned quantization method to perform Int8 quantization, and these quantized features (such as the second features or preset features) are saved as single bytes in the AI ​​camera storage to reduce camera storage usage. At the same time, the modulus of each row of quantized features (i.e., the feature modulus of each second feature or preset feature) is saved to complete the construction of the quantitative feature database, laying the foundation for subsequent direct use of the second features for matching.

[0080] In some embodiments, target ID information is obtained based on quantitative feature query matching; that is, based on the first type of query feature (such as the first feature), the target feature is determined in the quantitative feature database. This process mainly completes the matching of the query feature (such as the first feature) in the quantitative feature database to obtain the target ID information. The specific steps are as follows:

[0081] S31. Use the above-mentioned quantization method for performing Int8 quantization encoding on Float type features to perform Int8 quantization on the query feature (such as the first feature) to obtain the second feature.

[0082] S32: Use the second feature to directly perform feature matching with the quantitative feature database.

[0083] Figure 6 is a schematic diagram of a feature query process according to an embodiment of the present invention. As shown in Figure 6, the second feature is first matrix multiplied with the quantitative feature database to obtain N-dimensional data, and then the obtained product result is divided by the feature modulus of N preset features in the quantitative feature database to obtain the feature similarity between the second feature and all preset features in the quantitative feature database.

[0084] S33. Find the row number of the maximum value in the N-dimensional similarity data, which is the query target ID; that is, determine the preset feature with the largest feature similarity as the target feature, or determine the target ID information of the target feature in the quantitative feature database.

[0085] The technical solution provided in this application has the advantages of fast calculation and less storage compared to the feature matching algorithm based on Float type; it can be applied to AI cameras with less hardware resources and has the advantage of lower hardware cost; and this solution is compatible with many camera manufacturers and can be used with all supported quantitative AI cameras, and can be applied to security scenarios to perform face recognition on suspicious targets; or applied to community access control to quickly match the faces of people entering and leaving, and has a wide range of uses.

[0086] According to an embodiment of the present invention, an embodiment of a feature quantization device is also provided. It should be noted that the feature quantization device can be used to execute the feature quantization method in the embodiment of the present invention, and the feature quantization method in the embodiment of the present invention can be executed in the feature quantization device.

[0087] Figure 7 is a schematic diagram of a feature quantization device according to an embodiment of the present invention. As shown in Figure 7, the device may include: an acquisition module 72, used to acquire a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; a quantization module 74, used to quantize the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a determination module 76, used to determine a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0088] It should be noted that the acquisition module 72 in this embodiment can be used to execute step S102 in the embodiment of the present application, the quantification module 74 in this embodiment can be used to execute step S104 in the embodiment of the present application, and the determination module 76 in this embodiment can be used to execute step S106 in the embodiment of the present application. The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiments.

[0089] In an embodiment of the present invention, a first feature of a first type is obtained, wherein the first feature includes first feature data of multiple dimensions; the first feature is quantized using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a second feature of a second type is determined based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, and the second feature The data is determined based on the reconstructed feature data. The first type and the second type are different. The second feature data meets the numerical range required by the second type, thereby realizing the conversion of the first feature of the first type into the second feature of the second type. Since the second feature data in the second feature is determined based on the feature coding of a preset quantization dictionary, a second feature based on the coding form can be obtained. The volume of the second feature data in the coding form is smaller than that of the first feature in the data form. Therefore, in the feature processing process, the hardware resources occupied by the second feature in the coding form will be smaller than the hardware resources occupied by the first feature in the data form, thereby achieving the purpose of reducing the data volume of the feature data, realizing the technical effect of reducing the occupation of hardware resources by the feature data, and thus solving the technical problem of the large occupation of hardware resources by the feature data.

[0090] In some embodiments, the device also includes: an acquisition submodule, used to quantize the first feature using a preset quantization dictionary to obtain multiple training features of the first type before obtaining the reconstructed feature, wherein the training feature includes training feature data of multiple dimensions; a clustering submodule, used to cluster the multiple training features to obtain a preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; a generation submodule, used to assign continuous feature codes to the preset number of sample features and generate a preset quantization dictionary.

[0091] In some embodiments, the quantization module includes: a calculation unit, used to calculate the absolute value of the difference between the first target data and each sample target data to obtain the target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; a first determination unit, used to determine the target sample feature based on the sample target data with the smallest target data difference; a second determination unit, used to encode the feature of the target sample feature and determine it as the reconstructed feature data of the reconstructed feature in the target dimension.

[0092] In some embodiments, the determination module includes: a processing unit for decentralizing the reconstructed feature to obtain a second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is determined according to the numerical range required by the second type.

[0093] In some embodiments, the device also includes: a first calculation submodule, which is used to calculate the feature modulus of the second feature after determining the second feature of the second type based on the reconstructed feature, when the first feature is the original feature, wherein the original feature is used to indicate the construction of a quantitative feature database; and a construction module, which is used to construct a quantitative feature database based on the second feature and the feature modulus of the second feature.

[0094] In some embodiments, the device also includes: a second calculation submodule, which is used to calculate the feature similarity between the second feature and each preset feature in the quantitative feature database after determining the second feature of the second type based on the reconstructed feature, when the first feature is a query feature, wherein the query feature is used to query the target feature in the quantitative feature database; and a determination submodule, which is used to determine the preset feature with the largest feature similarity as the target feature.

[0095] In some embodiments, the second calculation submodule includes: an acquisition subunit, used to obtain multiple preset features and the feature modulus of each preset feature in a quantitative database (i.e., a quantitative feature database); a calculation subunit, used to calculate the second feature and complete matrix multiplication with the quantitative database to obtain a product result; a determination subunit, used to divide the product result by the feature modulus of each preset feature to determine the feature similarity between the second feature and each preset feature.

[0096] The embodiment of the present invention can provide a computer terminal, which can be any computer terminal device in a computer terminal group. In this embodiment, the computer terminal can also be replaced by a terminal device such as a mobile terminal.

[0097] In this embodiment, the computer terminal may be located in at least one network device among a plurality of network devices of a computer network.

[0098] In this embodiment, the above-mentioned computer terminal can execute the program code of the following steps in the feature quantization method: obtaining a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one to the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; determining a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0099] In some embodiments, Figure 8 is a block diagram of a computer terminal according to an embodiment of the present invention. As shown in Figure 8 , the computer terminal 80 may include: one or more (only one is shown in the figure) processors 82 and a memory 84.

[0100] The memory can be used to store software programs and modules, such as program instructions / modules corresponding to the feature quantification method and device in the embodiments of the present invention. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implementing the above-mentioned feature quantification method. The memory may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely located relative to the processor, and these remote memories may be connected to the terminal 80 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0101] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: obtain a first feature of the first type, wherein the first feature includes first feature data of multiple dimensions; quantize the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one to the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; determine a second feature of the second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0102] In some embodiments, the processor may also execute program code for the following steps: before using a preset quantization dictionary to quantize the first feature to obtain a reconstructed feature, obtaining multiple training features of the first type, wherein the training features include training feature data of multiple dimensions; clustering the multiple training features to obtain a preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; assigning continuous feature codes to the preset number of sample features to generate a preset quantization dictionary.

[0103] In some embodiments, the processor may also execute the program code of the following steps: calculating the absolute value of the difference between the first target data and each sample target data to obtain the target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; determining the target sample feature based on the sample target data with the smallest target data difference; encoding the feature of the target sample feature to determine it as the reconstructed feature data of the reconstructed feature in the target dimension.

[0104] In some embodiments, the above-mentioned processor can also execute the program code of the following steps: decentralized the reconstructed feature to obtain a second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is determined according to the numerical range required by the second type.

[0105] In some embodiments, the processor may further execute program code for the following steps: after determining a second feature of the second type based on the reconstructed feature, when the first feature is an original feature, calculating a feature modulus of the second feature, wherein the original feature is used to indicate construction of a quantitative feature database; and constructing a quantitative feature database based on the second feature and the feature modulus of the second feature.

[0106] In some embodiments, the processor may further execute program code for the following steps: after determining a second feature of the second type based on the reconstructed feature, when the first feature is a query feature, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database, wherein the query feature is used to query the target feature in the quantitative feature database; and determining the preset feature with the greatest feature similarity as the target feature.

[0107] In some embodiments, the processor may also execute the program code of the following steps: obtaining multiple preset features and the feature modulus of each preset feature in the quantitative database; calculating the matrix multiplication of the second feature and the quantitative database to obtain the product result; dividing the product result by the feature modulus of each preset feature to determine the feature similarity between the second feature and each preset feature.

[0108] According to an embodiment of the present invention, a feature quantization scheme is provided. In the embodiment of the present invention, a first feature of a first type is obtained, wherein the first feature includes first feature data of multiple dimensions; the first feature is quantized using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one with the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; a second feature of a second type is determined based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, and the second feature The data is determined based on the reconstructed feature data. The first type and the second type are different. The second feature data meets the numerical range required by the second type, thereby realizing the conversion of the first feature of the first type into the second feature of the second type. Since the second feature data in the second feature is determined based on the feature coding of a preset quantization dictionary, a second feature based on the coding form can be obtained. The volume of the second feature data in the coding form is smaller than that of the first feature in the data form. Therefore, in the feature processing process, the hardware resources occupied by the second feature in the coding form will be smaller than the hardware resources occupied by the first feature in the data form, thereby achieving the purpose of reducing the data volume of the feature data, realizing the technical effect of reducing the occupation of hardware resources by the feature data, and thus solving the technical problem of the large occupation of hardware resources by the feature data.

[0109] Those skilled in the art will appreciate that the structure shown in FIG8 is merely illustrative, and that the computer terminal may also be a smartphone (e.g., an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, a mobile internet device (MID), a PAD, or other terminal device. FIG8 does not limit the structure of the aforementioned electronic devices. For example, the computer terminal 80 may include more or fewer components (e.g., a network interface, a display device, etc.) than those shown in FIG8 , or may have a configuration different from that shown in FIG8 .

[0110] A person skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a non-volatile medium. The non-volatile storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0111] The embodiment of the present invention further provides a non-volatile storage medium. In this embodiment, the non-volatile storage medium can be used to store the program code executed by the feature quantization method provided in the above embodiment.

[0112] In this embodiment, the non-volatile storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0113] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: obtaining a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds one-to-one to the first feature data, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of the target dimension is the feature code of the target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; determining a second feature of a second type based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type and the second type are different, and the second feature data meets the numerical range required by the second type.

[0114] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: before quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, obtaining multiple training features of the first type, wherein the training features include training feature data of multiple dimensions; clustering the multiple training features to obtain a preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; assigning continuous feature codes to the preset number of sample features to generate a preset quantization dictionary.

[0115] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: calculating the absolute value of the difference between the first target data and each sample target data to obtain the target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; determining the target sample feature based on the sample target data with the smallest target data difference; encoding the feature of the target sample feature and determining it as the reconstructed feature data of the reconstructed feature in the target dimension.

[0116] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: decentralizing the reconstructed feature to obtain a second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is determined according to the numerical range required by the second type.

[0117] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: after determining a second feature of the second type based on the reconstructed feature, when the first feature is an original feature, calculating the feature modulus of the second feature, wherein the original feature is used to indicate the construction of a quantitative feature database; and constructing a quantitative feature database based on the second feature and the feature modulus of the second feature.

[0118] In this embodiment, the non-volatile storage medium is configured to store program code for performing the following steps: after determining a second feature of the second type based on the reconstructed feature, when the first feature is a query feature, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database, wherein the query feature is used to query the target feature in the quantitative feature database; and determining the preset feature with the largest feature similarity as the target feature.

[0119] In this embodiment, the non-volatile storage medium is configured to store program code for executing the following steps: obtaining multiple preset features and a feature modulus of each preset feature in a quantitative database; calculating a second feature and performing matrix multiplication with the quantitative database to obtain a product result; dividing the product result by the feature modulus of each preset feature to determine the feature similarity between the second feature and each preset feature.

[0120] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0121] In the above embodiments of the present invention, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0123] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0124] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0125] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a non-volatile storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned non-volatile storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, and other media that can store program code.

[0126] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A feature quantification method, comprising: Acquire a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds to the first feature data one-to-one, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of a target dimension is a feature code of a target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; A second feature of a second type is determined based on the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, the second feature data is determined based on the reconstructed feature data, the first type is different from the second type, and the second feature data meets the numerical range required by the second type.

2. According to the method of claim 1, before quantizing the first feature using a preset quantization dictionary to obtain a reconstructed feature, the method further comprises: Acquire a plurality of training features of the first type, wherein the training features include training feature data of multiple dimensions; Clustering the plurality of training features to obtain the preset number of sample features, wherein the preset number is determined according to the numerical range required by the second type; Continuous feature codes are assigned to the preset number of sample features to generate the preset quantization dictionary.

3. According to the method according to any one of claims 1 to 2, quantizing the first feature using a preset quantization dictionary to obtain the reconstructed feature comprises: Calculate the absolute value of the difference between the first target data and each sample target data to obtain the target data difference, wherein the first target data is the first feature data of the first feature in the target dimension, and the sample target data is the sample feature data of the sample feature in the target dimension; Determining target sample characteristics according to the sample target data having the smallest target data difference; The feature encoding of the target sample feature is determined as the reconstructed feature data of the reconstructed feature in the target dimension.

4. According to the method according to any one of claims 1 to 3, determining the second feature of the second type according to the reconstructed feature comprises: The reconstructed feature is decentralized to obtain the second feature, wherein the second feature data in the second feature is the difference between the reconstructed feature data of the same dimension and a preset constant, and the preset constant is calculated based on the second feature. The type is determined by the range of values ​​required.

5. The method according to any one of claims 1 to 4, after determining a second feature of a second type according to the reconstructed feature, the method further comprises: When the first feature is an original feature, calculating a feature modulus of the second feature, wherein the original feature is used to indicate the construction of a quantitative feature database; The quantitative feature database is constructed according to the second feature and the feature mode of the second feature.

6. The method according to any one of claims 1 to 5, after determining a second feature of a second type according to the reconstructed feature, the method further comprises: When the first feature is a query feature, calculating a feature similarity between the second feature and each preset feature in a quantitative feature database, wherein the query feature is used to query a target feature in the quantitative feature database; The preset feature with the greatest feature similarity is determined as the target feature.

7. The method according to claim 6, calculating the feature similarity between the second feature and each preset feature in the quantitative feature database comprises: Acquire a plurality of preset features and a feature mode of each of the preset features in a quantitative feature database; Calculate the second feature and the quantitative feature database to complete matrix multiplication to obtain a product result; The product result is divided by the feature modulus of each of the preset features to determine the feature similarity between the second feature and each of the preset features.

8. According to the method according to any one of claims 1 to 7, the preset number of sample features recorded in the preset quantization dictionary is determined according to the numerical range required by the second type.

9. The method according to any one of claims 1-8, wherein the first type comprises a floating point type, and the second type comprises an integer type.

10. A feature quantification device, comprising: An acquisition module, configured to acquire a first feature of a first type, wherein the first feature includes first feature data of multiple dimensions; a quantization module, configured to quantize the first feature using a preset quantization dictionary to obtain a reconstructed feature, wherein the preset quantization dictionary includes a preset number of sample features, each sample feature includes: a continuous feature code and sample feature data of multiple dimensions, the sample feature data of each dimension corresponds to the first feature data one-to-one, the reconstructed feature includes reconstructed feature data of multiple dimensions, the reconstructed feature data of a target dimension is a feature code of a target sample feature, and the target sample feature is the sample feature that is closest to the first feature data and the sample feature data in the target dimension; A determination module is configured to determine a second feature of a second type according to the reconstructed feature, wherein the second feature includes second feature data of multiple dimensions, and the second feature data is determined according to the reconstructed feature data. The first type is different from the second type, and the second characteristic data meets the value range required by the second type.

11. A non-volatile storage medium, the non-volatile storage medium being used to store a program, wherein: When the program is running, the device where the non-volatile storage medium is located is controlled to execute the feature quantification method described in any one of claims 1 to 9.

12. An electronic device comprising: A memory and a processor, wherein the processor is used to run a program stored in the processor, wherein the program executes the feature quantization method according to any one of claims 1 to 9 when running.

13. A computer program comprising: An instruction, when executed by a processor, causes the processor to perform the feature quantization method according to any one of claims 1-9.

Images