Compressed matrix splitting design method and device, equipment and storage medium

Abstract

The application discloses a compression matrix splitting design method, system, device and storage medium, and belongs to the field of image compression. First parameter matrix and second parameter matrix are generated by acquiring picture data; a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix; the process matrix is processed by a preset deep frequency domain neural network to generate correction data; a convergence value is acquired, and one of the following steps is executed according to the correction data and the convergence value: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by deep frequency domain neural network processing, and the data acquisition step is jumped to; or, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by Kronecker product operation to generate a compression matrix. A compression matrix design scheme which can be designed for different compression data is effectively provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image compression, and in particular to a compression matrix splitting design method, device, equipment and storage medium. BACKGROUND

[0002] In related technologies, when designing a compression matrix, a decisive matrix that has been designed is mostly directly used, but this method cannot be optimized according to changes in compression data. Therefore, how to provide a compression matrix splitting design method that can be optimized according to changes in compression data is a problem to be solved at present. SUMMARY

[0003] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides a compression matrix splitting design method, which can improve the design efficiency of the compression matrix.

[0004] The present application also provides a compression matrix splitting design device.

[0005] The present application also provides a compression matrix splitting design device.

[0006] The present application also provides a computer-readable storage medium.

[0007] The compression matrix splitting design method according to the first aspect of the present application comprises:

[0008] Obtaining picture data;

[0009] According to the picture data, a first parameter matrix and a second parameter matrix are generated;

[0010] Data acquisition step: acquiring the first parameter matrix and the second parameter matrix, and generating a process matrix according to the picture data, the first parameter matrix and the second parameter matrix;

[0011] Obtaining the process matrix and processing the process matrix through a preset deep frequency domain neural network to generate correction data;

[0012] Obtaining a convergence value, and performing one of the following steps according to the correction data and the convergence value:

[0013] If the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated through the deep frequency domain neural network processing, and the data acquisition step is jumped to;

[0014] Or,

[0015] If the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by a Kronecker product operation to generate a compressed matrix.

[0016] The compressed matrix splitting design method according to the embodiments of the present application has at least the following beneficial effects: the compressed matrix splitting design method provided by the present application can be used to obtain picture data and generate a first parameter matrix and a second parameter matrix according to the picture data. Through the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix, so that the process matrix is obtained. The process matrix is processed by a preset deep frequency domain neural network to generate correction data. According to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by the deep frequency domain neural network processing, and the data obtaining step is jumped to; or, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by a Kronecker product operation to generate a compressed matrix. In the design process of the compressed matrix, the first parameter matrix and the second parameter matrix are used to compress the picture. The correction data obtained by processing the compression result by the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix. A compressed matrix with good compression effect is obtained, and a compressed matrix design scheme that can be designed for different compression data is effectively provided.

[0017] According to some embodiments of the present application, the first parameter matrix and the second parameter matrix are generated according to the picture data, including:

[0018] The initial number of rows and the initial number of columns of the picture data are obtained.

[0019] A first parameter matrix with the same number of columns as the initial number of rows is generated according to the initial number of rows.

[0020] A second parameter matrix with the same number of rows as the initial number of columns is generated according to the initial number of columns.

[0021] According to some embodiments of the present application, the data obtaining step: obtaining the first parameter matrix and the second parameter matrix, and generating a process matrix according to the picture data, the first parameter matrix and the second parameter matrix, includes:

[0022] The second parameter matrix is multiplied by the picture data to obtain a calculation matrix.

[0023] The first parameter matrix is multiplied by the calculation matrix to generate the process matrix.

[0024] According to some embodiments of the present application, the process matrix is obtained, and the process matrix is processed by a preset deep frequency domain neural network to generate correction data, including:

[0025] The process matrix is processed by a preset deep frequency domain neural network to generate a processing result;

[0026] The processing result is obtained, and the processing result is processed by a full-link layer to generate error information;

[0027] The error information is obtained, and a classification operation is performed on the error information to generate correction data.

[0028] According to some embodiments of the present application, if the correction data is greater than or equal to the convergence value, the first parameter matrix and the second parameter matrix are generated by the deep frequency domain neural network processing, and the data acquisition step is jumped to, including:

[0029] If the correction data is greater than or equal to the convergence value, the parameter change matrix is generated by the reverse update operation of the deep frequency domain neural network;

[0030] The parameter change matrix, the current first parameter matrix and the current second parameter matrix are obtained, and the updated first parameter matrix and the updated second parameter matrix are generated according to the parameter change matrix, the current first parameter matrix and the current second parameter matrix, and the data acquisition step is jumped to.

[0031] According to some embodiments of the present application, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by a Kronecker product operation to generate a compression matrix, including:

[0032] The first parameter matrix and the second parameter matrix are obtained;

[0033] The second parameter matrix is transposed to obtain a third parameter matrix;

[0034] The third parameter matrix and the second parameter matrix are subjected to a Kronecker product operation to generate a compression matrix.

[0035] According to the compression matrix splitting design device of the second aspect of the embodiments of the present application, including:

[0036] The picture acquisition module is used for acquiring picture data;

[0037] The parameter matrix generation module is used for generating a first parameter matrix and a second parameter matrix according to the picture data;

[0038] a data acquisition module configured to perform a data acquisition step to acquire the first parameter matrix and the second parameter matrix, and generate a process matrix according to the picture data, the first parameter matrix and the second parameter matrix;

[0039] a deep frequency domain neural network processing module configured to acquire the process matrix, and process the process matrix through a preset deep frequency domain neural network to generate correction data

[0040] a judging module configured to acquire the correction data and a convergence value, and perform one of the following steps according to the correction data and the convergence value:

[0041] if the correction data is greater than or equal to the convergence value, generate an updated first parameter matrix and an updated second parameter matrix through the deep frequency domain neural network processing, and jump to the data acquisition step;

[0042] or,

[0043] if the correction data is less than the convergence value, calculate the current first parameter matrix and the current second parameter matrix through a Kronecker product operation to generate a compression matrix.

[0044] The compression matrix splitting design device according to the embodiment of the present application has at least the following beneficial effects: the picture acquisition module acquires picture data; the parameter matrix generation module can generate a first parameter matrix and a second parameter matrix according to the picture data; the data acquisition module can acquire the first parameter matrix and the second parameter matrix, and generate a process matrix according to the picture data, the first parameter matrix and the second parameter matrix; the deep frequency domain neural network processing module can acquire the process matrix, and process the process matrix through a preset deep frequency domain neural network to generate correction data; the judging module can acquire the correction data and a convergence value, and perform one of the following steps according to the correction data and the convergence value: if the correction data is greater than or equal to the convergence value, generate an updated first parameter matrix and an updated second parameter matrix through the deep frequency domain neural network processing, and jump to the data acquisition step; or, if the correction data is less than the convergence value, calculate the current first parameter matrix and the current second parameter matrix through a Kronecker product operation to generate a compression matrix. In the design process of the compression matrix, the first parameter matrix and the second parameter matrix are used to compress the picture, the correction data obtained by processing the compression result through the deep frequency domain neural network is used to continuously modify the first parameter matrix and the second parameter matrix, a compression matrix with good compression effect is obtained, and a compression matrix design scheme that can be used to design different compression data is effectively provided.

[0045] According to some embodiments of the present application, the parameter matrix generation module comprises:

[0046] a picture data analysis unit configured to obtain an initial row number and an initial column number of picture data;

[0047] a parameter matrix configuration unit configured to generate a first parameter matrix with the same column number as the initial row number according to the initial row number, and generate a second parameter matrix with the same row number as the initial column number according to the initial column number.

[0048] The compressed matrix splitting design device according to the third aspect of the present application comprises:

[0049] at least one processor, and

[0050] a memory in communication connection with the at least one processor; wherein

[0051] the memory stores instructions, and the instructions are executed by the at least one processor, so that the at least one processor implements the compressed matrix splitting design method according to the first aspect of the present application.

[0052] The compressed matrix splitting design device according to the embodiments of the present application has at least the following beneficial effects: the compressed matrix splitting design device according to the present application can implement the compressed matrix splitting design method according to the first aspect of the present application by storing the compressed matrix splitting design method according to the first aspect of the present application in the memory and using the at least one processor in communication connection with the memory, thereby obtaining picture data and generating a first parameter matrix and a second parameter matrix according to the picture data. By executing the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix. The process matrix is obtained, and the process matrix is processed by the preset deep frequency domain neural network to generate correction data. According to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by the deep frequency domain neural network processing and the data obtaining step is jumped to; or if the correction data is less than the convergence value, the current first parameter matrix and the current second parameter matrix are calculated by the Kronecker product operation to generate the compressed matrix. In the design process of the compressed matrix, the first parameter matrix and the second parameter matrix are used to compress the picture, the correction data obtained by processing the compression result by the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix, a compressed matrix with good compression effect is obtained, and an effective compressed matrix design scheme that can be designed for different compression data is provided.

[0053] The computer readable storage medium according to the fourth aspect of the present application comprises:

[0054] The computer readable storage medium stores computer executable instructions for performing the compressed matrix splitting design method according to the first aspect of the present application.

[0055] The computer readable storage medium according to the present application has at least the following beneficial effects: the computer readable storage medium can provide the compressed matrix splitting design method according to the first aspect of the present application stored therein, thereby obtaining picture data and generating a first parameter matrix and a second parameter matrix according to the picture data. By performing the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix; thereby obtaining the process matrix, and processing the process matrix through the preset deep frequency domain neural network to generate correction data, and according to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated through the deep frequency domain neural network processing and the data obtaining step is jumped to; or, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated through the Kronecker product operation to generate the compressed matrix. In the design process of the compressed matrix, the first parameter matrix and the second parameter matrix are used to compress the picture, the correction data obtained by processing the compression result through the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix, a compressed matrix with good compression effect is obtained, and a compressed matrix design scheme that can be designed for different compression data is effectively provided. Additional aspects and advantages of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood by practicing the present application. BRIEF DESCRIPTION OF DRAWINGS

[0056] The present application will be further described below in conjunction with the drawings and embodiments, wherein:

[0057] Figure 1 The flowchart of the compressed matrix splitting design method according to the present application.

[0058] Figure 2 The flowchart of step S200 in the embodiment. Figure 1

[0059] The flowchart of step S300 in the embodiment. Figure 3 Figure 1 The flowchart of step S300 in the embodiment.

[0060] Figure 4 The flowchart of step S300 in the embodiment.​Figure 1 The specific flowchart of step S400 is shown in the following.

[0061] Figure 5 For Figure 1 The specific flowchart of step S610 is shown in the following.

[0062] Figure 6 For Figure 1 The specific flowchart of step S620 is shown in the following.

[0063] Figure 7 The structural diagram of the compression matrix splitting design device of the embodiment of the present application is shown in the following.

[0064] Figure 8 For Figure 7 The specific structural diagram of the parameter matrix generation module 200 is shown in the following.

[0065] The figure mark: 100, picture acquisition module; 200, parameter matrix generation module; 300, data acquisition module; 400, deep frequency domain neural network processing module; 500, judgment module; 210, picture data analysis unit; 220, parameter matrix configuration unit. DETAILED DESCRIPTION

[0066] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary, only for explaining the present application, and cannot be understood as a limitation of the present application.

[0067] In the description of the present application, it is understood that the orientation description, such as the orientation or position relationship indicated by up, down, front, back, left, right, etc. is based on the orientation or position relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and cannot be understood as indicating or implying that the device or element indicated must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present application.

[0068] In the description of the present application, if the meaning of several is more than one, the meaning of multiple is more than two, greater than, less than, more than, etc. is not included in the number, above, below, etc. is understood to include the number. If it is described as first, second, only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the order of indicated technical features.

[0069] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be understood in a broad sense, and the person skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0070] In the description of the application, the description of the terms "one embodiment", "some embodiments", "exemplary embodiment", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the application. In the description, the exemplary description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0071] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application. In addition, the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0072] In the first aspect, with reference to Figure 1 In some embodiments, the application provides a design method of a compression matrix, comprising:

[0073] S100, acquiring picture data;

[0074] S200, generating a first parameter matrix and a second parameter matrix according to the picture data;

[0075] S300, data acquisition step: acquiring the first parameter matrix and the second parameter matrix, and generating a process matrix according to the picture data, the first parameter matrix and the second parameter matrix;

[0076] S400, acquiring the process matrix, and processing the process matrix through a preset deep frequency domain neural network to generate correction data;

[0077] S500, acquiring a convergence value,

[0078] and executing one of the following steps according to the correction data and the convergence value:

[0079] S610, if the correction data is greater than or equal to the convergence value, generating an updated first parameter matrix and an updated second parameter matrix through deep frequency domain neural network processing and jumping to the data acquisition step;

[0080] or,

[0081] S620, if the correction data is less than the convergence value, calculating the current first parameter matrix and the second parameter matrix through Kronecker product operation to generate a compression matrix.

[0082] The compression matrix splitting design method provided by the present application can realize obtaining picture data, and generating a first parameter matrix and a second parameter matrix according to the picture data. By performing the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix; thus the process matrix is obtained, and the process matrix is processed by the preset deep frequency domain neural network to generate correction data, and according to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by the deep frequency domain neural network processing and the data obtaining step is jumped to; or, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by the Kronecker product operation to generate the compression matrix. In the design process of the compression matrix, the first parameter matrix and the second parameter matrix are used to realize the compression of the picture, the correction data obtained by processing the compression result by the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix, the compression matrix with good compression effect is obtained, and an effective compression matrix design scheme which can be designed for different compression data is provided. The convergence value is not fixed, and different setting values are adopted for different compression images and different parameter setting deep frequency domain neural networks.

[0083] In addition, the compression matrix splitting design method provided by the present application uses the Kronecker product operation to optimize the matrix parameters of the first parameter matrix and the second parameter matrix, greatly reduces the calculation amount, significantly improves the optimization speed of each element of the compression matrix, and thus provides an efficient compression matrix splitting design method.

[0084] In some specific embodiments, for a general matrix with 100 rows and 100 columns, when the general matrix is not split by the Kronecker product, 10000 elements need to be calculated, by splitting the general matrix into a first parameter matrix and a second parameter matrix with 10 rows and 10 columns, the present application only needs to calculate 200 elements in the design process of the compression matrix, which significantly reduces the calculation amount and effectively improves the design efficiency of the compression matrix.

[0085] Reference Figure 2 In some embodiments, step S200 includes:

[0086] S210, obtaining an initial row number and an initial column number of picture data;

[0087] S220, generating a first parameter matrix with the same column number as the initial row number according to the initial row number;

[0088] S230, generating a second parameter matrix with the same number of rows and the same number of columns as the initial number of columns according to the initial number of rows.

[0089] By obtaining the initial number of rows and the initial number of columns of the picture data, and determining the number of columns of the first parameter matrix and the number of rows of the second parameter matrix according to the initial number of rows and the initial number of columns. This method can simplify the elements contained in the compression matrix, reduce the amount of data required in the process of calculating the compression matrix, thereby reducing the calculation time of the compression process.

[0090] In some specific embodiments, the number of rows of the first parameter matrix is less than the number of columns of the first parameter matrix, and the number of columns of the second parameter matrix is less than the number of rows of the second parameter matrix, so as to ensure that the matrix size of the picture data processed by the first parameter matrix and the second parameter matrix is smaller than the original picture data size, thereby realizing the compression processing of the picture data.

[0091] In some embodiments, referring to Figure 3 , step S300 includes:

[0092] S310, performing matrix multiplication on the second parameter matrix and the picture data to obtain a calculation matrix.

[0093] S320, performing matrix multiplication on the first parameter matrix and the calculation matrix to generate a process matrix.

[0094] By obtaining the second parameter matrix and the picture data, and performing matrix multiplication on the second parameter matrix as the left matrix and the picture data, a calculation matrix can be obtained; wherein the calculation matrix contains the picture data after preliminary compression. According to the calculation matrix and the first parameter matrix, the first parameter matrix is used as the right matrix of the calculation matrix to perform matrix multiplication, thereby obtaining a process matrix. The process matrix is a compressed file after compression processing of the picture data. This method can obtain the first parameter matrix and the second parameter matrix generated by splitting the general matrix, and perform matrix right multiplication and matrix left multiplication on the picture data according to the first parameter matrix and the second parameter matrix, thereby realizing the compression processing of the picture data, effectively reducing the operation amount and improving the compression efficiency.

[0095] In some embodiments, referring to Figure 4 , step S400 includes:

[0096] S410, processing the process matrix by a preset deep frequency domain neural network to generate a processing result;

[0097] S420, obtaining the processing result, processing the processing result by a full connection layer to generate error information;

[0098] S430, obtaining the error information and performing classification operation on the error information to generate correction data.

[0099] The present application realizes the pre-extraction of error information by inputting the process matrix containing picture data into the depth frequency domain neural network for training to obtain a processing result containing all error information, and realizes the extraction of error information in the compression matrix splitting design process by obtaining the processing result containing all error information, performing full-linkage layer calculation and processing on the error information in the processing result to obtain summarized error information, performing classification operation on the summarized information, and generating correction data containing error information.

[0100] In some other embodiments, the depth frequency domain neural network includes at least one group of convolution layer, pooling layer and activation function. The convolution layer can perform zero value padding on the convolution kernel therein according to the number of rows and columns of the process matrix to obtain a first intermediate matrix, wherein the number of rows of the first intermediate matrix is the same as the number of rows of the process matrix, and the number of columns of the first intermediate matrix is the same as the number of columns of the process matrix; and the first intermediate matrix is subjected to discrete cosine transform processing to obtain a second intermediate matrix; then the process matrix and the second intermediate matrix are subjected to matrix point multiplication processing to generate first processing data; the pooling layer can perform pooling processing on the first processing data to generate second processing data; and the second processing data is subjected to inverse discrete cosine transform processing; and the activation function can perform activation processing on the second processing data to generate the corresponding processing result. The processing result here is only the processing result after one depth frequency domain neural network processing, and in some specific embodiments, the processing result after one depth frequency domain neural network processing can be input again into the preset depth frequency domain neural network for processing to reduce the error caused by the large data error due to the accidental factors in the obtained data. Through the convolution layer, the pooling layer and the activation function processing, compared with the traditional neural network, there are fewer layers and the calculation process is more efficient. The preset activation function can be a Relu function, a Sigmoid function, etc., without limitation.

[0101] In some embodiments, S500 includes obtaining a convergence value

[0102] The compression matrix splitting design method provided by the application, by obtaining the convergence value, and according to the size relationship between the convergence value and the correction data, one of the execution steps S610 and S620 is judged, the compression matrix designed for different pictures represented by different correction data is generated, thereby effectively providing a compression matrix design scheme which can be designed for different compression data. The size of the convergence value is determined according to general experience and does not have a fixed requirement.

[0103] With reference to Figure 5 In some embodiments, step S610 comprises:

[0104] S611, if the correction data is greater than or equal to the convergence value, the reverse update operation is performed through the deep frequency domain neural network to generate the parameter change matrix.

[0105] S612, the parameter change matrix, the current first parameter matrix and the current second parameter matrix are obtained, and the updated first parameter matrix and the updated second parameter matrix are generated according to the parameter change matrix, the current first parameter matrix and the current second parameter matrix, and the data acquisition step is jumped to.

[0106] The application obtains the correction data, and compares the correction data with the convergence value. When the comparison result is that the correction data is greater than or equal to the convergence value, the deep frequency domain neural network is updated in reverse according to the correction data, and the corresponding parameter change matrix is obtained. According to the obtained parameter change matrix, the current first parameter matrix and the current second parameter matrix are subjected to matrix addition operation with the parameter change matrix, so as to update the corresponding first parameter matrix and second parameter matrix, and the updated general matrix is submitted to the data acquisition step to provide the updated first parameter matrix and second parameter matrix, and the compression matrix splitting design method provided by the application is executed. The compression matrix splitting design method provided by the application provides a means for updating the first parameter matrix and the second parameter matrix according to the identification result, a method for simplifying the compression matrix splitting design process, and improves the efficiency of the compression matrix splitting design.

[0107] With reference to Figure 6 In some embodiments, S620 comprises:

[0108] S621, the first parameter matrix and the second parameter matrix are obtained.

[0109] S622, the second parameter matrix is transposed to obtain the third parameter matrix.

[0110] S623, the third parameter matrix and the second parameter matrix are subjected to Kronecker product operation to generate the compression matrix.

[0111] The application obtains the current first parameter matrix and the current second parameter matrix, and performs transpose processing on the current second parameter matrix, and performs Kronecker product operation on the second parameter matrix after the transpose processing and the first parameter matrix, to obtain the compression matrix corresponding to the current first parameter matrix and the second parameter matrix. The picture data is compressed by the first parameter matrix and the second parameter matrix, so that the calculation amount can be reduced. In addition, the first parameter matrix and the second parameter matrix are obtained by Kronecker product operation, and the compression matrix corresponding to the first parameter matrix and the second parameter matrix can be obtained, so that the difficulty of compression matrix splitting design is reduced, and the design efficiency of the compression matrix is improved.

[0112] In a second aspect, referring to Figure 7 The application provides a compression matrix splitting design device, comprising:

[0113] The picture acquisition module 100 is configured to acquire picture data.

[0114] The parameter matrix generation module 200 is configured to generate a first parameter matrix and a second parameter matrix according to the picture data.

[0115] The data acquisition module 300 is configured to acquire the first parameter matrix and the second parameter matrix, and generate a process matrix according to the picture data, the first parameter matrix and the second parameter matrix.

[0116] The deep frequency domain neural network processing module 400 is configured to acquire the process matrix, and process the process matrix by a preset deep frequency domain neural network to generate correction data.

[0117] The judgment module 500 is configured to acquire the correction data and a convergence value, and perform one of the following steps according to the correction data and the convergence value: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by the deep frequency domain neural network processing, and the data acquisition step is jumped to; or if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by the Kronecker product operation to generate the compression matrix.

[0118] The compression matrix splitting design device provided by the present invention obtains image data through the image acquisition module 100; through the parameter matrix generation module 200, a first parameter matrix and a second parameter matrix can be generated according to the image data; through the data acquisition module 300, the first parameter matrix and the second parameter matrix can be obtained, and a process matrix can be generated according to the image data, the first parameter matrix and the second parameter matrix; through the deep frequency domain neural network processing module 400, the process matrix can be obtained, and the process matrix can be processed by a preset deep frequency domain neural network to generate correction data; through the judgment module 500, the correction data and the convergence value can be obtained, and one of the following steps can be performed according to the correction data and the convergence value: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated through deep frequency domain neural network processing and the process is skipped to the data acquisition step; or, if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated through the Kronecker product operation to generate a compression matrix. In the design process of the compression matrix, the image is compressed by the first parameter matrix and the second parameter matrix. The compression result is processed by the deep frequency domain neural network 400, and the correction data obtained is continuously corrected to the first parameter matrix and the second parameter matrix, so as to obtain a compression matrix with better compression effect. This effectively provides a compression matrix design scheme that can be targeted for different compression data.

[0119] Reference Figure 8 In some embodiments, the parameter matrix generation module 200 includes:

[0120] The picture data analysis unit 210 is used to obtain the initial number of rows and columns of the picture data;

[0121] The parameter matrix configuration unit 220 is used to generate a first parameter matrix with the same number of columns as the initial number of rows according to the initial number of rows; and to generate a second parameter matrix with the same number of rows as the initial number of columns according to the initial number of columns.

[0122] The compression matrix splitting design device can obtain the initial row number and the initial column number of the picture data through the picture data analysis unit 210, and can generate a first parameter matrix with the same column number and the initial row number according to the initial row number through the parameter matrix configuration unit 220, and generate a second parameter matrix with the same row number and the initial column number according to the initial column number. The elements in the first parameter matrix and the second parameter matrix generated according to the initial row number and the initial column number can be preset by the system or completely random, and are not limited. The picture data is compressed by the obtained first parameter matrix and the second parameter matrix, which can effectively simplify the compression process, that is, reduce the calculation amount required in the process matrix generation process, reduce the design difficulty of the compression matrix, and improve the compression matrix splitting design efficiency.

[0123] In some specific embodiments, the parameter matrix configuration unit 220 generates a first parameter matrix with a row number less than a column number, and generates a second parameter matrix with a column number less than a row number, so that the picture data is compressed by the first parameter matrix and the second parameter matrix generated by the parameter matrix configuration unit 220 to generate compressed data smaller than the original picture data, thereby realizing the compression processing of the picture data.

[0124] In a third aspect, the present application provides a compression matrix splitting design device, comprising:

[0125] at least one processor, and

[0126] a memory connected in communication with the at least one processor; wherein

[0127] The memory stores instructions, and the instructions are executed by the at least one processor, so that the at least one processor implements the compression matrix splitting design method according to the first aspect of the present application when executing the instructions.

[0128] The compression matrix splitting design device provided by the present application can realize the compression matrix splitting design method in the first aspect of the present application, and can realize obtaining picture data, and generating a first parameter matrix and a second parameter matrix according to the picture data. By executing the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix; thereby obtaining the process matrix, and processing the process matrix through the preset deep frequency domain neural network to generate correction data, and according to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated through the deep frequency domain neural network processing, and the data obtaining step is jumped to; or, if the correction data is less than the convergence value, the current first parameter matrix and the current second parameter matrix are calculated through the Kronecker product operation to generate the compression matrix. In the design process of the compression matrix, the first parameter matrix and the second parameter matrix are used to realize the compression of the picture, the correction data obtained by processing the compression result by the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix, the compression matrix with good compression effect is obtained, and an effective compression matrix design scheme which can be designed for different compression data is provided. In the fourth aspect, the present application provides a computer readable storage medium, and the computer readable storage medium stores computer executable instructions, which can be used to execute the compression matrix splitting design method in the first aspect of the present application.

[0129] According to the computer readable storage medium of the embodiment of the present application, the compressed matrix splitting design method stored in the computer readable storage medium can be executed. According to the computer readable storage medium of the embodiment of the present application, at least the following beneficial effects are achieved: the computer readable storage medium can provide the compressed matrix splitting design method of the first aspect of the present application stored therein to obtain picture data and generate a first parameter matrix and a second parameter matrix according to the picture data. By executing the data obtaining step, the first parameter matrix and the second parameter matrix are obtained, and a process matrix is generated according to the picture data, the first parameter matrix and the second parameter matrix. Thus, the process matrix is obtained, and the process matrix is processed by the preset deep frequency domain neural network to generate correction data. According to the size relationship between the correction data and the convergence value, the following operations are performed: if the correction data is greater than or equal to the convergence value, the updated first parameter matrix and the updated second parameter matrix are generated by the deep frequency domain neural network processing, and the data obtaining step is jumped to; or if the correction data is less than the convergence value, the current first parameter matrix and the second parameter matrix are calculated by the Kronecker product operation to generate the compressed matrix. In the design process of the compressed matrix, the picture is compressed by the first parameter matrix and the second parameter matrix, the correction data obtained by processing the compression result by the deep frequency domain neural network is used to continuously correct the first parameter matrix and the second parameter matrix, the compressed matrix with good compression effect is obtained, and an effective compressed matrix design scheme that can be designed for different compression data is provided.

Claims

1. A compression matrix splitting design method, characterized in that: include: Get image data; Obtaining the initial number of rows and columns of the image data; According to the initial number of rows, generating a first parameter matrix with the same number of columns as the initial number of rows; generating, according to the initial number of columns, a second parameter matrix having the same number of rows as the initial number of columns; Data acquisition step: performing matrix multiplication on the second parameter matrix and the image data to obtain a calculation matrix; Performing matrix multiplication on the first parameter matrix and the calculation matrix to generate a process matrix; Obtaining the process matrix, and processing the process matrix through a preset deep frequency domain neural network to generate correction data; Obtain a convergence value, and perform one of the following steps according to the correction data and the convergence value: If the corrected data is greater than or equal to the convergence value, a reverse update operation is performed through the deep frequency domain neural network to generate a parameter change matrix; the parameter change matrix and the current first parameter matrix and the current second parameter matrix are obtained, and an updated first parameter matrix and an updated second parameter matrix are generated according to the parameter change matrix, the current first parameter matrix and the current second parameter matrix, and the process jumps to the data acquisition step; or, If the corrected data is less than the convergence value, obtaining the first parameter matrix and the second parameter matrix; The second parameter matrix is ​​transposed to obtain a third parameter matrix; and a Kronecker product operation is performed on the third parameter matrix and the first parameter matrix to generate a compressed matrix.

2. The compression matrix splitting design method according to claim 1, characterized in that: The obtaining of the process matrix and processing the process matrix through a preset deep frequency domain neural network to generate correction data includes: Processing the process matrix through a preset deep frequency domain neural network to generate a processing result; Obtaining the processing result, processing the processing result through a full link layer, and generating error information; The error information is acquired, and a classification operation is performed on the error information to generate correction data.

3. A compressed matrix splitting design device, characterized in that: include: Image acquisition module, used to obtain image data; A picture data analysis unit, configured to obtain an initial number of rows and columns of the picture data; a parameter matrix configuration unit, the parameter matrix configuration unit being configured to generate, according to the initial number of rows, a first parameter matrix having the same number of columns as the initial number of rows; and generating a second parameter matrix having the same number of rows as the initial number of columns according to the initial number of columns; a data acquisition module, configured to execute a data acquisition step to perform matrix multiplication on the second parameter matrix and the image data to obtain a calculation matrix; Performing matrix multiplication on the first parameter matrix and the calculation matrix to generate a process matrix; A deep frequency domain neural network processing module is used to obtain the process matrix and process the process matrix through a preset deep frequency domain neural network to generate correction data; A judgment module is configured to obtain the correction data and the convergence value, and perform one of the following steps according to the correction data and the convergence value: If the corrected data is greater than or equal to the convergence value, a reverse update operation is performed through the deep frequency domain neural network to generate a parameter change matrix; the parameter change matrix and the current first parameter matrix and the current second parameter matrix are obtained, and an updated first parameter matrix and an updated second parameter matrix are generated according to the parameter change matrix, the current first parameter matrix and the current second parameter matrix, and the process jumps to the data acquisition step; or, If the corrected data is less than the convergence value, obtaining the first parameter matrix and the second parameter matrix; The second parameter matrix is ​​transposed to obtain a third parameter matrix; and a Kronecker product operation is performed on the third parameter matrix and the first parameter matrix to generate a compressed matrix.

4. A compressed matrix splitting design device, characterized in that: include: at least one processor, and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions, and the instructions are executed by the at least one processor, so that when the at least one processor executes the instructions, the compressed matrix splitting design method according to any one of claims 1 to 2 is implemented.

5. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which can be used to execute the compressed matrix splitting design method according to any one of claims 1 to 2.

Images