Image feature extraction and privacy protection identification method

Abstract

The application provides an image feature extraction and privacy protection identification method, comprising the following steps: encoding different components of a privacy image into a quaternion matrix; performing multi-resolution singular value decomposition in the form of quaternions on the quaternion matrix to obtain a first approximate component, and performing multi-resolution singular value decomposition in the form of quaternions on the first approximate component to obtain a second approximate component; performing sparse random projection on the quaternion matrix, the first approximate component and the second approximate component; performing feature extraction on the results of sparse random projection through a concatenated quaternion two-dimensional discrete cosine transform network; and inputting the extracted features into a classifier for identification. Through the method provided by the application, multi-view or multi-modal image components can be effectively fused and the revocability can be met, the security is ensured, the reliable identification precision is ensured, and the method can be applied to the fields of multimedia information security and visual content protection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and pattern recognition technology. Background Technology

[0002] With the widespread adoption of mobile smart terminals and the mobile internet, massive amounts of multimedia data are generated, transmitted, stored, and shared via the open internet. Digital images may contain confidential or private information, such as important details like geographical location and spectral data of ground features in remote sensing images, and personal health information, related medical procedures, and costs in medical images. While the sharing and transmission of large amounts of digital images via the network offers convenience, it also poses risks such as information theft and malicious tampering. If private information in images is leaked during network transmission, the consequences could extend beyond personal financial losses to include personal injury and even national security threats. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] Therefore, the purpose of this invention is to propose an image feature extraction and privacy protection identification method for fusing image components from multiple perspectives or multiple modalities.

[0005] To achieve the above objectives, a first aspect of the present invention provides an image feature extraction and privacy protection identification method, comprising:

[0006] Encode the different components of the privacy image into a quaternion matrix;

[0007] The quaternion matrix is ​​subjected to multi-resolution singular value decomposition in quaternion form to obtain a first approximate component, and the first approximate component is subjected to multi-resolution singular value decomposition in quaternion form to obtain a second approximate component.

[0008] Sparse random projection is performed on the quaternion matrix, the first approximation component, and the second approximation component;

[0009] The results of sparse random projection are used for feature extraction through a cascaded quaternion two-dimensional discrete cosine transform network.

[0010] The extracted features are input into a classifier for recognition.

[0011] In addition, the image feature extraction and privacy protection identification method according to the above embodiments of the present invention may also have the following additional technical features:

[0012] Furthermore, in one embodiment of the present invention, encoding the different components of the privacy image into a quaternion matrix includes:

[0013] The different components of the privacy image are represented as real and imaginary components, respectively, as a quaternion matrix f. q (x,y), with dimensions N×M, where (x,y) represents the spatial coordinates of an image pixel.

[0014] Furthermore, in one embodiment of the present invention, the step of performing multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain the first approximate component includes:

[0015] The quaternion matrix f q (x, y) is divided into non-overlapping 2×2 sub-blocks and adjusted to 4×1 column vectors; by stacking the columns, a matrix X of size 4×NM / 4 is obtained. q ;

[0016] Calculate X q The corresponding decentralized matrix

[0017] according to Calculate the scattering matrix Where H represents the conjugate transpose;

[0018] The eigenvector matrix U1 is obtained by performing quaternion singular value decomposition on the scattering matrix T1;

[0019] Reconstructed matrix Will The first line is adjusted to The first approximate component is obtained.

[0020] Furthermore, in one embodiment of the present invention, the sparse random projection onto the quaternion matrix, the first approximation component, and the second approximation component includes:

[0021] For matrix Sparse random projection, expressed by the formula:

[0022]

[0023]

[0024] Furthermore, in one embodiment of the present invention, feature extraction is performed on the results of sparse random projection using a cascaded quaternion two-dimensional discrete cosine transform network, including:

[0025] For matrix Features were extracted using cascaded quaternion two-dimensional discrete cosine transform networks, and the three feature vectors were concatenated.

[0026] Furthermore, in one embodiment of the present invention, the matrix is... The process of feature extraction using a cascaded quaternion two-dimensional discrete cosine transform network includes:

[0027] Two-dimensional discrete cosine transform bases of size k×k, L1 and L2, are selected as filters and sorted in a zigzag pattern according to the main order of horizontal frequencies to obtain filter banks W1 and W2.

[0028] In the first stage, for each quaternion matrix Two-dimensional convolution is performed using L1 filter banks W1 of size k×k, and Tanh is used as the activation function to perform a nonlinear mapping on the convolution result:

[0029]

[0030] Where * represents a two-dimensional convolution with zero-padding, used to keep the matrix size N×M unchanged before and after convolution;

[0031] In the second stage, L2 filter banks W2 of size k×k are used to process the feature matrix. Perform two-dimensional convolution to output a feature matrix of size N×M.

[0032]

[0033] In the binary hash encoding stage, the matrix The real and imaginary components are encoded separately and divided into B blocks to extract histograms. The histograms of all sub-blocks are concatenated and normalized. The normalized features are combined into a quaternion matrix as the face features.

[0034] To achieve the above objectives, a second aspect of the present invention provides an image feature extraction and privacy protection identification device, comprising the following modules:

[0035] The encoding module is used to encode the different components of the privacy image into a quaternion matrix;

[0036] The decomposition module is used to perform multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain a first approximate component, and then perform multi-resolution singular value decomposition in quaternion form on the first approximate component to obtain a second approximate component.

[0037] The projection module is used to perform sparse random projection on the quaternion matrix, the first approximation component, and the second approximation component;

[0038] The extraction module is used to extract features from the results of sparse random projection through a cascaded quaternion two-dimensional discrete cosine transform network.

[0039] The recognition module is used to input the extracted features into the classifier for recognition.

[0040] To achieve the above objectives, a third aspect of the present invention provides a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements an image feature extraction and privacy protection identification method as described above.

[0041] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements an image feature extraction and privacy protection identification method as described above.

[0042] The image feature extraction and privacy-preserving recognition method proposed in this invention recognizes images based on multi-resolution singular value decomposition (SVD) in the quaternion domain and a two-dimensional discrete cosine transform (DCT) network. This aims to solve the problems of complex network structure, numerous network layers, and long training time in existing convolutional neural networks. Specifically, the technical solution involves: encoding different image components into quaternion matrices; performing quaternion multi-resolution singular value decomposition; applying sparse random projection to the decomposition results; extracting features through a cascaded quaternion DCT network; and inputting the features into a classifier for training and recognition. This invention can effectively fuse multi-view or multi-modal image components while satisfying revocability, ensuring reliable recognition accuracy while maintaining security. It can be applied to the fields of multimedia information security and visual content protection. Attached Figure Description

[0043] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0044] Figure 1 This is a schematic flowchart of an image feature extraction and privacy protection identification method provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of an image feature extraction and privacy protection identification device provided in an embodiment of the present invention. Detailed Implementation

[0046] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] The image feature extraction and privacy protection identification method of the present invention is described below with reference to the accompanying drawings.

[0048] Figure 1 This is a schematic flowchart of an image feature extraction and privacy protection identification method provided in an embodiment of the present invention.

[0049] like Figure 1 As shown, the image feature extraction and privacy protection identification method includes the following steps:

[0050] S101: Encode the different components of the privacy image into a quaternion matrix;

[0051] S102: Perform multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain the first approximate component, and then perform multi-resolution singular value decomposition in quaternion form on the first approximate component to obtain the second approximate component.

[0052] S103: Perform sparse random projection on the quaternion matrix, the first approximation component, and the second approximation component;

[0053] S104: The results of sparse random projection are used to extract features through a cascaded quaternion two-dimensional discrete cosine transform network;

[0054] S105: Input the extracted features into the classifier for recognition.

[0055] Furthermore, in one embodiment of the present invention, encoding different components of the privacy image into a quaternion matrix includes:

[0056] The different components of the privacy image are represented as real and imaginary components, respectively, as a quaternion matrix f. q (x,y), with dimensions N×M, where (x,y) represents the spatial coordinates of an image pixel.

[0057] Furthermore, in one embodiment of the present invention, performing multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain a first approximate component includes:

[0058] The quaternion matrix f q (x, y) is divided into non-overlapping 2×2 sub-blocks and adjusted to 4×1 column vectors; by stacking the columns, a matrix X of size 4×NM / 4 is obtained. q ;

[0059] Calculate X q The corresponding decentralized matrix

[0060] according to Calculate the scattering matrix Where H represents the conjugate transpose;

[0061] The eigenvector matrix U1 is obtained by performing quaternion singular value decomposition on the scattering matrix T1;

[0062] Reconstructed matrix Will The first line is adjusted to The first approximate component is obtained.

[0063] For matrix Then, perform the aforementioned quaternion multi-resolution singular value decomposition to obtain approximate components.

[0064] Furthermore, in one embodiment of the present invention, sparse random projection is performed on the quaternion matrix, the first approximation component, and the second approximation component, including:

[0065] For matrix Sparse random projection, expressed by the formula:

[0066]

[0067]

[0068] Furthermore, in one embodiment of the present invention, feature extraction is performed on the results of sparse random projection using a cascaded quaternion two-dimensional discrete cosine transform network, including:

[0069] For matrix Features were extracted using cascaded quaternion two-dimensional discrete cosine transform networks, and the three feature vectors were concatenated.

[0070] Furthermore, in one embodiment of the present invention, the matrix is... The process of feature extraction using a cascaded quaternion two-dimensional discrete cosine transform network includes:

[0071] Two-dimensional discrete cosine transform bases of size k×k, L1 and L2, are selected as filters and sorted in a zigzag pattern according to the main order of horizontal frequencies to obtain filter banks W1 and W2.

[0072] In the first stage, for each quaternion matrix Two-dimensional convolution is performed using L1 filter banks W1 of size k×k, and Tanh is used as the activation function to perform a nonlinear mapping on the convolution result:

[0073]

[0074] Where * represents a two-dimensional convolution with zero-padding, used to keep the matrix size N×M unchanged before and after convolution;

[0075] In the second stage, L2 filter banks W2 of size k×k are used to process the feature matrix. Perform two-dimensional convolution to output a feature matrix of size N×M.

[0076]

[0077] In the binary hash encoding stage, the matrix The real and imaginary components are encoded separately and divided into B blocks to extract histograms. The histograms of all sub-blocks are concatenated and normalized. The normalized features are combined into a quaternion matrix as the face features.

[0078] QSoftmax is used as a classifier to train and recognize the extracted features.

[0079] The image feature extraction and privacy-preserving recognition method proposed in this invention recognizes images based on multi-resolution singular value decomposition (SVD) in the quaternion domain and a two-dimensional discrete cosine transform (DCT) network. This aims to solve the problems of complex network structure, numerous network layers, and long training time in existing convolutional neural networks. Specifically, the technical solution involves: encoding different image components into quaternion matrices; performing quaternion multi-resolution singular value decomposition; applying sparse random projection to the decomposition results; extracting features through a cascaded quaternion DCT network; and inputting the features into a classifier for training and recognition. This invention can effectively fuse multi-view or multi-modal image components while satisfying revocability, ensuring reliable recognition accuracy while maintaining security. It can be applied to the fields of multimedia information security and visual content protection.

[0080] Advantages of the present invention's method compared to existing technologies:

[0081] (1) Quaternion multi-resolution singular value decomposition is used to extract image components at different scales, which avoids complex calculations compared to wavelet decomposition.

[0082] (2) Compared with convolutional neural networks, quaternion discrete cosine transform networks have higher model efficiency.

[0083] (3) The visual content protection of the image is achieved by using sparse matrix projection, which has security.

[0084] To achieve the above embodiments, the present invention also proposes an image feature extraction and privacy protection identification device.

[0085] Figure 2 This is a schematic diagram of an image feature extraction and privacy protection identification device provided in an embodiment of the present invention.

[0086] like Figure 2As shown, the image feature extraction and privacy protection recognition device includes: an encoding module 100, a decomposition module 200, a projection module 300, an extraction module 400, and a recognition module 500, wherein...

[0087] The encoding module is used to encode the different components of the privacy image into a quaternion matrix;

[0088] The decomposition module is used to perform multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain the first approximate component, and then perform multi-resolution singular value decomposition in quaternion form on the first approximate component to obtain the second approximate component.

[0089] The projection module is used to perform sparse random projections on the quaternion matrix, the first approximation component, and the second approximation component.

[0090] The extraction module is used to extract features from the results of sparse random projection through a cascaded quaternion two-dimensional discrete cosine transform network.

[0091] The recognition module is used to input the extracted features into the classifier for recognition.

[0092] To achieve the above objectives, a third aspect of the present invention provides a computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the image feature extraction and privacy protection identification method as described above.

[0093] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the image feature extraction and privacy protection identification method as described above.

[0094] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0095] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0096] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An image feature extraction and privacy protection identification method, characterized in that, The method comprises the following steps: The different components of the privacy image are encoded as quaternion matrices, where the different components of the privacy image are represented as real and imaginary components, respectively, in quaternion matrices. The size is N × M ,in( x , y () represents the spatial coordinates of an image pixel; performing multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain a first approximation component, and performing multi-resolution singular value decomposition in quaternion form on the first approximation component to obtain a second approximation component; performing sparse random projection on the quaternion matrix, the first approximation component, and the second approximation component; performing feature extraction on the result of sparse random projection through a cascaded quaternion two-dimensional discrete cosine transform network; inputting the extracted features into a classifier for recognition; the performing of the multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain the first approximation component comprises: The quaternion matrix is divided into non-overlapping 2x2 sub-blocks and adjusted to be 4x1 column vectors; by stacking the columns, a matrix of size 4x NM / 4 is obtained ; Computations Corresponding decentralized matrix ; According to Computing the scattering matrix where H denotes the conjugate transpose; On scattering matrix Performing quaternion singular value decomposition to obtain eigenvector matrix ; reconstruction matrix : , the first row of is adjusted to to obtain a first approximation component .

2. The method of claim 1, wherein, the performing of the sparse random projection on the quaternion matrix, the first approximation component, and the second approximation component comprises: Sparse random projections are applied to the matrix , , } and can be expressed by the formula ， 。 3. The method of claim 2, wherein, the performing of the feature extraction on the result of sparse random projection through the cascaded quaternion two-dimensional discrete cosine transform network comprises: On the matrix The features are extracted using cascaded quaternion two-dimensional discrete cosine transform networks respectively, and the three feature vectors are spliced.

4. The method of claim 3, wherein, The matrix The process of feature extraction by a cascaded quaternion two-dimensional discrete cosine transform network includes: k k ;​​​​​​ In the first stage, for each quaternion matrix use Each size is k × k filter bank Perform a two-dimensional convolution and use Tanh as the activation function to apply a non-linear mapping to the convolution result: ， wherein, is a two-dimensional convolution with zero padding operation for keeping the matrix size unchanged before and after the convolution; is a two-dimensional convolution with zero padding operation for keeping the matrix size unchanged before and after the convolution; In the second stage, a 2D convolution is performed on the feature matrix using a filter bank of size k × k , resulting in a feature matrix of size N × M ,​​​​ ， In the binary hash coding stage, the real and imaginary parts of the matrix are encoded respectively, divided into B blocks, and the histograms of all sub-blocks are extracted; the histograms of all sub-blocks are spliced together and normalized, and the normalized features are combined into a quaternion matrix as the face features.

5. An image feature extraction and privacy protection identification device, characterized in that, The method comprises the following modules: The encoding module is used to encode the different components of the privacy image into quaternion matrices, where the different components of the privacy image are represented as real and imaginary components, respectively, in quaternion matrices. The size is N × M ,in( x , y () represents the spatial coordinates of an image pixel; a decomposition module configured to perform multi-resolution singular value decomposition in quaternion form on the quaternion matrix to obtain a first approximation component, and perform multi-resolution singular value decomposition in quaternion form on the first approximation component to obtain a second approximation component; a projection module configured to perform sparse random projection on the quaternion matrix, the first approximation component, and the second approximation component; an extraction module configured to perform feature extraction on the result of sparse random projection through a cascaded quaternion two-dimensional discrete cosine transform network; a recognition module configured to input the extracted features into a classifier for recognition; the decomposition module is further configured to: The quaternion matrix is partitioned into non-overlapping 2x2 sub-blocks and adjusted to 4x1 column vectors; by stacking the columns, a matrix of size 4x NM / 4 is obtained ; Computations Corresponding decentralized matrix ; According to Computing the scattering matrix where H denotes the conjugate transpose; On a scattering matrix Performing a quaternion singular value decomposition to obtain an eigenvector matrix ; reconstruction matrix : , the first row of is adjusted to to obtain a first approximation component .

6. A computer device, comprising: The computer program is executed by the processor to implement the image feature extraction and privacy protection recognition method according to any one of claims 1-4.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the image feature extraction and privacy protection recognition method according to any one of claims 1-4.

Images