Processing and recognition of image data, method and apparatus for training an image recognition model

Abstract

Embodiments of the present specification provide a method and apparatus for processing and recognizing image data, and training an image recognition model. When processing image data, the image data is subjected to frequency domain transformation-based data desensitization processing to obtain desensitized image data of the image data, the obtained desensitized image data comprising a set of subgraph data, each subgraph data in the set of subgraph data corresponding to a different frequency. Subsequently, the obtained desensitized image data is subjected to data enhancement-based image mixing processing to obtain data-enhanced desensitized image data. Optionally, each subgraph data in the data-enhanced desensitized image data is subjected to graph size alignment processing, such that each subgraph data after the graph size alignment processing has the same size as the original image data.

Description

Technical Field

[0001] The embodiments in this specification generally relate to the field of artificial intelligence technology, and in particular to image data processing methods, image recognition model training methods, image recognition methods and apparatuses thereof. Background Technology

[0002] Image-based data processing solutions have been widely applied in numerous applications, such as facial recognition-based payment systems. These solutions require raw image data as input to third-party devices for processing. Due to the unpredictable security of the operating environment of these third-party devices, the raw image data, or the feature vector data obtained after feature extraction, is susceptible to malicious theft, posing a significant threat to data security or business security (e.g., secure payments). Therefore, before providing raw image data to third-party devices, it is necessary to utilize provided image desensitization processing solutions to protect data privacy. Summary of the Invention

[0003] In view of the above, embodiments of this specification provide an image data processing method, an image recognition model training method, an image recognition method, and an apparatus thereof. Using this image data processing method and apparatus, desensitized image data is obtained by performing frequency domain transform-based data desensitization processing on the image data, and data augmentation-based image mixing processing is performed on the obtained desensitized image data, thereby achieving data privacy protection for the original image data.

[0004] According to one aspect of the embodiments of this specification, a method for image data processing is provided, comprising: performing data desensitization processing on image data based on frequency domain transformation to obtain desensitized image data of the image data, the desensitized image data including a set of sub-image data, each sub-image data in the set of sub-image data corresponding to a different frequency; and performing image mixing processing on the sub-image data in the obtained desensitized image data based on data enhancement to obtain data-enhanced desensitized image data.

[0005] Alternatively, in one example of the foregoing aspects, the data augmentation may include one of the following data augmentations: Mixup data augmentation and Instahide data augmentation.

[0006] Optionally, in one example of the above aspects, the data augmentation includes Mixup data augmentation, where the number of images used for mixing is k, and the maximum weight of the mixed images is W. maxThe image blending process based on data augmentation for the obtained desensitized image data may include: shuffling the sub-image data set of the desensitized image data k-1 times to obtain k sub-image data sets; constructing an image hypermatrix of size m*k based on the obtained k sub-image data sets, where the first column of the image hypermatrix corresponds to the original sub-image data set, and m is the number of sub-image data sets in the original sub-image data set; randomly generating weight coefficients for each image in the image hypermatrix; and performing row normalization on the weight coefficients of the images in the image hypermatrix so that the sum of the weight coefficients of each row of images is 1, and the weight coefficient of each image is no greater than W. max ; and perform a weighted summation on each row of the supermatrix to obtain a hybrid image supermatrix of size m*1, wherein the images in the hybrid image supermatrix are desensitized image data after data augmentation.

[0007] Alternatively, in one example of the above aspects, the W max The value of is 0.55, and the value of k is 3.

[0008] Optionally, in one example of the above aspects, the method may further include: performing graph size alignment processing on each sub-graph data in the data-enhanced desensitized image data, such that the size of each sub-graph data after graph size alignment processing is the same as the size of the original image data.

[0009] Optionally, in one example of the above aspects, data desensitization processing of image data based on frequency domain transformation may include: performing local frequency domain transformation processing on the image data to obtain at least one feature map, each feature map including multiple elements and corresponding to a data block in the image data, each element corresponding to a frequency in the frequency domain; constructing frequency component channel feature maps corresponding to each frequency using the elements corresponding to each frequency in the at least one feature map respectively; and selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps to obtain desensitized image data of the image data, wherein the selected target frequency component channel feature map is a key channel feature for image recognition.

[0010] Optionally, in one example of the above aspects, after filtering out at least one target frequency component channel feature map from the constructed frequency component channel feature map, the method may further include: performing a first shuffling process on the target frequency component channel feature map to obtain a first shuffling feature map; and performing a normalization process on the first shuffling feature map to obtain desensitized image data of the image data.

[0011] Optionally, in one example of the above aspects, selecting at least one target frequency component channel feature map from the constructed frequency component channel feature map may include: selecting at least one target frequency component channel feature map from the constructed frequency component channel feature map based on channel importance or based on a preset filtering rule.

[0012] Optionally, in one example of the above aspects, after normalizing the first shuffle feature map, the method may further include: performing channel mixing on the normalized first shuffle feature map; performing a second shuffle on the channel-mixed first shuffle feature map to obtain a second shuffle feature map; and normalizing the second shuffle feature map to obtain desensitized image data of the image data.

[0013] Optionally, in one example of the above aspects, the number of target frequency component channel feature maps is one more than the feature dimension of the first feature data. Channel mixing of the first shuffle feature map may include: channel mixing of two adjacent frequency component channel feature maps in the first shuffle feature map.

[0014] Alternatively, in one example of the above aspects, the image data includes face image data.

[0015] According to another aspect of the embodiments of this specification, an apparatus for image data processing is provided, comprising: a data desensitization processing unit, which performs data desensitization processing on image data based on frequency domain transformation to obtain desensitized image data of the image data, the desensitized image data including a set of sub-image data, each sub-image data in the set of sub-image data corresponding to a different frequency; and an image mixing processing unit, which performs image mixing processing on the sub-image data in the obtained desensitized image data based on data enhancement to obtain data-enhanced desensitized image data.

[0016] According to another aspect of the embodiments of this specification, a training method for an image recognition model is provided. The training method includes: acquiring image data and its labels as training samples; processing the image data using the method described above to obtain data-enhanced de-identified image data of the image data, the data-enhanced de-identified image data including a set of sub-image data, each sub-image data having a weight coefficient; performing label mixing processing on the corresponding labels of each sub-image data of each de-identified image data using the weight coefficients of each sub-image data of each de-identified image data; and training an image recognition model using the each sub-image data of the de-identified image data and its label-mixed labels.

[0017] Optionally, in one example of the above aspects, when the sub-image data involved in the image blending process comes from the same category, the labels of each sub-image data involved in the image blending process are kept unchanged, and when at least some of the sub-image data involved in the image blending process comes from different categories, the non-zero values ​​in the labels of each sub-image data are adjusted to k non-zero values, each non-zero value corresponding to a weight coefficient of a sub-image data.

[0018] Optionally, in one example of the above aspects, the image data includes face image data, and the label or the recognition result indicating the identity information corresponding to the face. The image recognition model is used to predict the feature vector representation of the image data based on the de-identified image data of the image data, and to determine the recognition result based on the predicted feature vector representation.

[0019] According to another aspect of the embodiments of this specification, a training apparatus for an image recognition model is provided, comprising: a training data acquisition unit for acquiring image data and its labels serving as training samples; an image processing unit for processing the image data using the method described above to obtain data-enhanced desensitized image data of the face image data, the data-enhanced desensitized image data including a set of sub-image data, each sub-image data having a weight coefficient, and using the weight coefficients of each sub-image data of each desensitized image data to perform label mixing processing on the corresponding labels of each sub-image data of each desensitized image data; and a model training unit for training an image recognition model using the each sub-image data of the desensitized image data and its labels after label mixing processing.

[0020] According to another aspect of the embodiments of this specification, an image recognition method is provided, comprising: acquiring image data to be recognized; processing the image data using the method described above to obtain data-enhanced de-identified image data of the image data; providing the de-identified image data to an image recognition model to predict feature vector representations of the image data; and determining whether the image data to be recognized contains the target object based on the predicted feature vector representations and feature vector representations of the target object.

[0021] Optionally, in one example of the above aspects, the image data to be identified includes facial image data, and the target object includes facial image data with specific identity information.

[0022] According to another aspect of the embodiments of this specification, an image recognition apparatus is provided, comprising: an image data acquisition unit for acquiring image data to be recognized; an image data processing unit for processing the image data using the method described above to obtain data-enhanced de-identified image data of the image data; a feature vector representation prediction unit for providing the de-identified image data to an image recognition model to predict the feature vector representation of the image data; and an image recognition unit for determining whether the image data to be recognized contains the target object based on the predicted feature vector representation and the feature vector representation of the target object.

[0023] According to another aspect of the embodiments of this specification, an image processing apparatus is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the method for image data processing as described above.

[0024] According to another aspect of the embodiments of this specification, a training apparatus for an image recognition model is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the training method for the image recognition model as described above.

[0025] According to another aspect of the embodiments of this specification, an image recognition apparatus is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the image recognition method as described above.

[0026] According to another aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores executable instructions, which, when executed, cause a processor to perform the method for image data processing as described above, or to perform the method for training an image recognition model as described above, or to perform the image recognition method as described above.

[0027] According to another aspect of the embodiments of this specification, a computer program product is provided, including a computer program that is executed by a processor to implement the method for image data processing as described above, or the method for training an image recognition model as described above, or the image recognition method as described above. Attached Figure Description

[0028] A further understanding of the nature and advantages of this specification can be achieved by referring to the following figures. In the figures, similar components or features may have the same reference numerals.

[0029] Figure 1 An example schematic diagram of an application scenario of an image data processing system according to an embodiment of this specification is shown.

[0030] Figure 2 An example flowchart of an image data processing method according to an embodiment of this specification is shown.

[0031] Figure 3 An example flowchart of a frequency domain transformation-based data desensitization process according to an embodiment of this specification is shown.

[0032] Figure 4 An example schematic diagram illustrating the conversion of image feature data from the spatial domain to the frequency domain according to an embodiment of this specification is shown.

[0033] Figure 5 A schematic diagram illustrating an example of a local frequency domain transformation process according to an embodiment of this specification is shown.

[0034] Figure 6 An example schematic diagram of a frequency component channel feature map according to an embodiment of this specification is shown.

[0035] Figure 7 Another example flowchart of a frequency domain transform-based data desensitization process according to an embodiment of this specification is shown.

[0036] Figure 8 An example flowchart of an image blending process based on Mixup data enhancement according to an embodiment of this specification is shown.

[0037] Figure 9 An example structural diagram of an image recognition model according to an embodiment of this specification is shown.

[0038] Figure 10 An example flowchart of an image recognition model training method according to an embodiment of this specification is shown.

[0039] Figure 11 An example flowchart of an image recognition method according to an embodiment of this specification is shown.

[0040] Figure 12 An example block diagram of an image data processing apparatus according to an embodiment of this specification is shown.

[0041] Figure 13 An example block diagram of a data desensitization processing unit according to an embodiment of this specification is shown.

[0042] Figure 14 An example block diagram of a data desensitization processing unit according to another embodiment of this specification is shown.

[0043] Figure 15 An example block diagram of an image blending processing unit according to an embodiment of this specification is shown.

[0044] Figure 16 An example block diagram of an image recognition model training apparatus according to an embodiment of this specification is shown.

[0045] Figure 17 An example block diagram of an image recognition device according to an embodiment of this specification is shown.

[0046] Figure 18 An example schematic diagram of an image data processing apparatus based on a computer system according to an embodiment of this specification is shown.

[0047] Figure 19 An example schematic diagram of an image recognition model training apparatus based on a computer system, according to an embodiment of this specification, is shown.

[0048] Figure 20 An example schematic diagram of an image recognition device based on a computer system according to an embodiment of this specification is shown. Detailed Implementation

[0049] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0050] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0051] With the development of computer technology, image recognition is increasingly penetrating all aspects of people's lives. For example, facial recognition technology has been widely applied in various scenarios, such as smart locks, terminal application logins, and facial payment. However, facial images may contain sensitive information or personal privacy information. To avoid unnecessary losses caused by the leakage of this information, it is necessary to de-identify facial images.

[0052] Given that facial images are a crucial basis for identity verification, one implementation of facial recognition technology involves inputting facial images into a facial recognition model for processing, and then using the processing results (e.g., facial feature data) for identity verification. However, due to the uncontrollable security of third-party devices (or their operating environments), the original facial image data, or the facial feature data obtained after processing the original facial image data using a facial recognition model, is susceptible to malicious theft. Criminals may use the stolen facial feature data to reconstruct the original facial image data through reverse engineering, and then illegally obtain the user's access permissions based on the original facial image data, such as unlocking doors or making unauthorized transactions. This poses a significant threat to secure payments and unlocking systems. With increasing emphasis on personal privacy protection, how to protect the privacy of facial images used for facial recognition is an urgent issue that needs to be addressed.

[0053] Therefore, the embodiments of this specification propose an image processing method, an image recognition model training method, an image recognition method and apparatus, which protect the data privacy of image data more securely and effectively by performing desensitization processing on image data.

[0054] The examples above are for illustrative purposes only and should not be construed as limiting the scope of this solution. Although this specification primarily uses facial images as examples, it should be understood that the technical solutions disclosed herein can be applied to any type of image data requiring privacy protection, such as fingerprint image data. The technical solutions disclosed herein are illustrated below through a description of the accompanying drawings.

[0055] The following description, with reference to the accompanying drawings, describes an image processing method, an image recognition model training method, and an image recognition method and apparatus according to embodiments of this specification.

[0056] Figure 1 An example schematic diagram of an application scenario of an image data processing system 100 according to an embodiment of this specification is shown.

[0057] like Figure 1 As shown, in the application scenario, the image data processing system 100 may include a server 110, a network 120, a terminal device 130, and a storage device 140.

[0058] The image data processing system 100 can be widely applied to various image recognition scenarios, such as face unlocking, face payment, and face terminal application login. In some embodiments, it can also be applied to any other scenario requiring image privacy protection, such as the transmission and storage of sensitive image data. After processing image data using the methods disclosed in this specification, subsequent image data processing tasks (e.g., image recognition) can be completed normally based on the processed image data, and it is difficult to deduce the original image data from the processed image data, thereby effectively protecting the privacy information in the image from being leaked.

[0059] In a typical application scenario, terminal device 130 can acquire a face image to be identified using an image acquisition device (such as a camera). Terminal device 130 can then perform de-identification processing on the acquired face image using the image processing method proposed in this specification to obtain de-identified image data. This de-identified image data is then sent to server 110 via network 120. Server 110 can be used to process information and / or data related to data service requests and / or image processing and image recognition. For example, in response to a data service request from terminal device 130, server 110 can receive the de-identified image data sent by terminal device 130, and after completing the de-identified image data recognition (e.g., determining that the de-identified image data comes from a legitimate face image), it can provide feedback on the recognition result to terminal device 130 or provide corresponding data services to terminal device 130. In some embodiments, server 110 can process the image using a pre-trained image recognition model to obtain a predicted feature vector representation of the face image. After obtaining the predicted feature vector representation of the face image, the server 110 can perform further operations, such as comparing it with the feature vector representation of the face image that has been successfully registered and stored in the system, and feeding back the comparison result (e.g., identity recognition result) to the terminal device 130, thereby completing face payment, face unlocking, etc.

[0060] In some embodiments, server 110 may be a local server or a remote server. For example, server 110 may locally connect to terminal device 130 to obtain information and / or data sent by terminal device 130. As another example, server 110 may remotely receive information and / or data sent by terminal device 130 via network 120. In some embodiments, server 110 may be implemented on a cloud platform. By way of example only, the cloud platform may include private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-tiered cloud, etc., or any combination thereof.

[0061] Network 120 can facilitate the exchange of information and / or data. In some embodiments, one or more components of the image data processing system 100 (e.g., server 110, terminal device 130, storage device 140) can transmit information to other components of the image data processing system 100 via network 120. For example, terminal device 130 can send image desensitization data of an image to be identified to server 120 via network 120. As another example, server 110 can send the predicted feature vector representation obtained by processing the image desensitization data of the identified image to storage device 140 for storage, and return the comparison result of the predicted feature vector representation and the feature vector representation of the target object to terminal device 130, etc. In some embodiments, network 120 can be any form of wired or wireless network, or any combination thereof. By way of example only, network 120 can be one or more combinations of wired networks, fiber optic networks, telecommunications networks, intranets, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), Bluetooth networks, etc.

[0062] Terminal device 130 can be used to process information and / or data associated with image processing and image recognition to perform one or more functions disclosed in this specification. In some embodiments, terminal device 130 may be a public device that provides image acquisition and / or data processing services to the public, such as Internet of Things (IoT) device 130-1. Exemplary IoT device 130-1 may include, but is not limited to, face-sensing vending machines, face-sensing payment devices, bank personal banking devices, and any combination thereof. After a user completes face recognition on terminal device 130, they can use the data services provided by that device. In some embodiments, terminal device 130 may be used to acquire image data to be processed acquired due to the triggering of the device's image acquisition device. In some embodiments, terminal device 130 may perform image data processing on the image data to be processed to obtain de-identified image data of the image data to be processed. In some embodiments, a trusted execution environment may be deployed on terminal device 130, and image acquisition and image processing may be performed within the trusted execution environment. In some embodiments, terminal device 130 may include one or more processing engines (e.g., a single-core processing engine or a multi-core processor). By way of example only, a processing engine may include one or more combinations of a central processing unit (CPU), an application-specific integrated circuit (ASIC), a special-purpose instruction set processor (ASIP), a graphics processing unit (GPU), a physical processing unit (PPU), a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic device (PLD), a controller, a microcontroller unit, a reduced instruction set computer (RISC), a microprocessor, etc.

[0063] In some embodiments, the terminal device 130 may also be a portable device with data acquisition, storage, and / or transmission functions, such as a tablet computer, laptop computer, smartphone, camera, or any combination thereof. In some embodiments, the terminal device 130 may interact with the server 110 via a network. For example, the terminal device 130 may transmit processed and de-identified image data to the server 110. In some embodiments, the data acquired by the terminal device 130 may be facial image data captured by the device's camera. The server 110 may receive the facial image data from the terminal device 130, perform de-identification processing on it, and then perform subsequent recognition. In this case, the server 110 may be integrated with the terminal device 130.

[0064] Storage device 140 can store data and / or instructions related to image processing, such as feature vectors, image data, and identity information of images of users who have successfully registered with the system. In some embodiments, storage device 140 can store data obtained / acquired by terminal device 130 and / or server 110. In some embodiments, storage device 140 can store data and / or instructions used by server 110 to perform or use in order to accomplish the exemplary methods described in this application. In some embodiments, storage device 140 may include mass storage, removable storage, volatile read-write storage, read-only storage (ROM), etc., or any combination thereof. Exemplary mass storage may include disks, optical disks, solid-state drives, etc. Exemplary removable storage may include flash drives, floppy disks, optical disks, memory cards, compact disks, magnetic tapes, etc. Exemplary volatile read-only storage may include random access memory (RAM). Exemplary RAM may include dynamic RAM (DRAM), double-rate synchronous dynamic RAM (DDRSDRAM), static RAM (SRAM), thyristor RAM (T-RAM), and zero-capacitance RAM (Z-RAM), etc. Exemplary ROMs may include mask ROMs (MROMs), programmable ROMs (PROMs), erasable programmable ROMs (PEROMs), electronically erasable programmable ROMs (EEPROMs), optical disc ROMs (CD-ROMs), and digital universal disk ROMs, etc. In some embodiments, the storage device 140 may be implemented on a cloud platform. By way of example only, the cloud platform may include private clouds, public clouds, hybrid clouds, community clouds, distributed clouds, internal clouds, multi-tiered clouds, etc., or any combination thereof.

[0065] In some embodiments, storage device 140 may be connected to network 120 to communicate with one or more components of image data processing system 100 (e.g., server 110, user terminal 130). One or more components of image data processing system 100 may access data or instructions stored in storage device 140 via network 120. In some embodiments, storage device 140 may be directly connected to or communicate with one or more components of image data processing system 100 (e.g., server 110, user terminal 130, etc.). In some embodiments, storage device 140 may be part of server 110.

[0066] Figure 2 An example flowchart of an image data processing method 200 according to an embodiment of this specification is shown. In some embodiments, the image data processing method 200 may be executed by a processing device, such as a terminal device 130 or a server 110. For example, the image data processing method 200 may be stored in a storage device (such as the built-in storage unit of the processing device or an external storage device) in the form of a program or instructions, which, when executed, can implement the image data processing method 200.

[0067] like Figure 2 As shown in step 210, the image data to be processed undergoes a frequency domain transformation-based data desensitization process to obtain desensitized image data. The obtained desensitized image data includes a set of sub-image data, where each sub-image data corresponds to a different frequency.

[0068] In this specification, the term "subimage data" can refer to multiple subimages extracted from image data to be processed using certain image processing techniques, each subimage carrying some features of the image data to be processed. In some cases, the term "subimage data" can be used interchangeably with the term "feature map".

[0069] Figure 3 An example flowchart of a frequency domain transformation-based data desensitization process 300 according to an embodiment of this specification is shown.

[0070] like Figure 3 As shown, in step 310, a local frequency domain transformation is performed on the image data to be processed to obtain at least one feature map. Each feature map includes multiple elements and corresponds to a data block in the image data. Each element corresponds to a frequency in the frequency domain.

[0071] Image data to be processed refers to image data that has not undergone anonymization. This data can be raw face image data or image data after face detection or face alignment has been performed on raw face image data. Raw face image data can refer to image data directly captured by an image acquisition device (e.g., a camera) without any processing. Face detection refers to detecting the position of a face in an image. The image data to be processed can be an image cropped based on the position of the face in the image, for example, by subtracting excess parts of the image other than the face. Face alignment refers to correcting the angle of the face in the image. The face in the original face image may be tilted at a certain angle; face alignment can straighten the face in the image to facilitate subsequent image recognition and processing.

[0072] In some embodiments, the processing device may acquire image data to be processed through the camera of the terminal device, or by reading from a database, storage device, or calling a data interface.

[0073] It should be noted that the program / code for acquiring the image data to be processed can run in a trusted execution environment deployed in the processing device. The security features of this execution environment ensure that the image data acquired by the processing device will not be stolen. Furthermore, the methods and / or processes disclosed in the embodiments of this description can also be executed in a trusted execution environment to ensure that the entire process from the source of image data acquisition to image data processing is secure and trustworthy, thereby improving the security of privacy protection for the image data to be processed.

[0074] After obtaining the image data to be processed as described above, a local frequency domain transformation can be performed on the obtained image data to obtain at least one feature map. A feature map refers to multiple sub-maps extracted from the image data to be processed using certain image processing techniques. Each sub-map carries some features of the image data to be processed. The obtained feature map can be the same size as the image data to be processed (e.g., a one-to-one correspondence of pixels), or it can be different in size from the image data to be processed.

[0075] In some embodiments, examples of local frequency domain transformation processing may include, but are not limited to: local discrete cosine transform, local wavelet transform, or local discrete Fourier transform.

[0076] Figure 4 A schematic diagram illustrating an example of converting speech feature data from the spatial domain to the frequency domain according to an embodiment of this specification is shown. Figure 4 In this context, the spatial domain is represented by the coordinate system (x, y), the frequency domain by the coordinate system (c, v), and N*M represents the image size. For example, Figure 4The value is 2*2. The number of feature points after spatial and frequency domain transformation can be the same. In the spatial domain, one square represents a pixel position, and in the frequency domain, one square represents a frequency position.

[0077] In some embodiments, the discrete cosine transform formula (1) shown below can be used to perform discrete cosine transform on the grayscale image data to be transformed.

[0078]

[0079] Where F(u,v) is the value of the feature point (i.e., each frequency position) in the converted frequency domain, f(x,y) is the pixel value in the image data (grayscale image data) to be transformed, (u,v) is the coordinate of the feature point in the converted frequency domain, (x,y) is the coordinate of the image data to be transformed in the spatial domain, N is the number of rows of pixels or feature points in the image data to be transformed, and M is the number of columns of pixels or feature points in the image data to be transformed. For example, when the size of the image data to be processed is 8*8, N=M=8.

[0080] c(u) can be represented by the following formula (2).

[0081]

[0082] Where c(u) = c(v).

[0083] By performing local frequency domain transformation on image data, multiple transformation results, i.e., multiple feature maps, can be obtained. During local frequency domain transformation, an image block (local image block) smaller than the image data to be transformed can be selected. For example, if the image data to be transformed is 256×256, the selected image block is 8×8. Then, the selected image block is used to move and sample the image data to be transformed at a certain step size (e.g., 8). The local data of the image data to be transformed (i.e., the 8×8 data block) obtained from each sampling is subjected to discrete cosine transform based on formulas (1) and (2), thus obtaining multiple transformation results, each of which can be 8×8 in size. The smaller the moving step size of the image block during discrete cosine transform, the more features the resulting transformation contains, which can help improve the accuracy of subsequent image data processing.

[0084] Figure 5 A schematic diagram illustrating an example of a local frequency domain transformation process according to an embodiment of this specification is shown. Figure 5In the example, the image data to be transformed is 6×6 in size, the selected local image patch is 2×2 in size, and the image data to be transformed is sampled with a step size of 2. A frequency domain transformation, such as a discrete cosine transform, is then performed on the sampled local image patch, resulting in 9 transformation results, i.e., 9 feature maps, namely 51, 52, 53, 54, 55, 56, 57, 58, and 59. The values ​​of each frequency position in each transformation result are represented by fi1, fi2, fi3, and fi4, where i represents the i-th transformation result, and fij represents the value of the j-th frequency position in the i-th transformation result. As can be seen from the figure, each transformation result has 4 corresponding frequency positions.

[0085] In step 320, elements corresponding to each frequency in at least one feature map are used to construct frequency component channel feature maps for each frequency. Specifically, elements (values) at the same frequency position in each transformation result are combined to obtain a frequency component channel feature map, thereby obtaining multiple frequency component channel feature maps at different frequency positions in the corresponding transformation results. It is easy to understand that the number of frequency component channel feature maps is consistent with the number of pixels in the image patch used for sampling during the transformation process.

[0086] Figure 6 An example schematic diagram of a frequency component channel feature map according to an embodiment of this specification is shown. Figure 6 The frequency component channel feature map shown in the figure is related to Figure 5 The frequency component channel feature map corresponding to the transformation result in the image.

[0087] like Figure 6 As shown, based on Figure 5 The transformation results in four frequency component channel feature maps (each image block used for sampling has four pixels), and each frequency component channel feature map includes nine elements. Frequency component channel feature map 61 corresponds to the first frequency position fi1, frequency component channel feature map 62 corresponds to the second frequency position fi2, frequency component channel feature map 63 corresponds to the third frequency position fi3, and frequency component channel feature map 64 corresponds to the fourth frequency position fi4.

[0088] After constructing the frequency component channel feature map as described above, at step 330, at least one target frequency component channel feature map is selected from the constructed frequency component channel feature map. The selected target frequency component channel feature map is the key channel feature for image recognition.

[0089] In some embodiments, at least one target frequency component channel feature map can be selected from the constructed frequency component channel feature map based on channel importance or based on preset filtering rules.

[0090] In some embodiments, the processing device can input multiple transformation results into a trained SENet network, which then provides the channel importance (e.g., a score positively correlated with importance) of each feature map. Here, channel importance refers to channel importance relative to image recognition. The SENet network can be trained together with the image recognition model (i.e., as a whole), for example, by adding the SENet network to the image recognition model and adjusting the parameters of the SENet network during the training of the image recognition model, thereby obtaining the SENet network used to determine the channel importance of the feature maps.

[0091] In some embodiments, the preset selection rule may be to select and retain a preset proportion of feature maps containing more feature information. For example, among multiple feature maps obtained after discrete cosine transform and reconstruction, a preset proportion of low-frequency feature maps may be selected to be retained, while some high-frequency feature maps may be discarded. For example, 50%, 60%, or 70% of the low-frequency feature maps may be retained, while the remaining high-frequency feature maps may be discarded. For example, retaining... Figure 6 The low-frequency feature maps 61, 62, and 63 shown are used, while the high-frequency feature map 64 is discarded. In the discrete cosine transform result, the values ​​at the upper left frequency position correspond to low-frequency components, and the values ​​at the lower right frequency position correspond to high-frequency components. For example... Figure 5 In the transformation result 51, f11 corresponds to low-frequency data, and f14 corresponds to high-frequency data. Refer to formula (1) above; when (u, v) is (0, 0). F(0,0) contains no AC component and can be considered as DC. Therefore, the frequency corresponding to the value in the upper left corner of the transformation result is the lowest. As the coordinate position moves to the lower right corner, F(u,v) contains AC component and the frequency increases. Therefore, the frequency corresponding to the value in the lower right corner of the transformation result is the highest.

[0092] exist Figure 3 In the example, at least one target frequency component channel feature map selected from the screening is used as the desensitized image data of the image data to be processed. In this way, the desensitized image data is image data after desensitizing multiple feature maps, and may include one or more frequency component channel feature maps after desensitization. The desensitized image data is different from the image data to be processed. Because the desensitized image data is a feature map, the original face information of the image data cannot be directly obtained from the desensitized image data.

[0093] exist Figure 3 In the example, the desensitization method includes reconstructing the frequency component channel feature map of the feature map and filtering the feature map of the frequency component channel feature map. In other embodiments, in addition to the above operations, the desensitization method may also include shuffling, normalization, channel mixing, or any combination thereof.

[0094] Figure 7 Another example flowchart of a frequency domain transform-based data desensitization process 700 according to an embodiment of this specification is shown. Figure 7 The illustrated embodiments are for Figure 3 Modifications to the illustrated embodiment. Figure 7 Steps 710 to 730 in the middle Figure 2 210 to 230 are exactly the same. For simplicity, the identical parts will not be repeated below, and only the differences will be described.

[0095] like Figure 7 As shown, after the target frequency component channel feature map is selected in 730, the selected target frequency component channel feature map is not used as desensitized image data. Instead, the operations from 740 to 760 are continued to obtain more secure desensitized image data.

[0096] Specifically, after selecting at least one target frequency component channel feature map from the constructed frequency component channel feature map, at 740, the selected target frequency component channel feature map is subjected to a first shuffle process to obtain a first shuffled feature map; and the first shuffled feature map is then normalized. In some embodiments, the data after normalizing the first shuffled feature map can be directly used as desensitized image data. In some embodiments, subsequent processing needs to be performed on the data after normalizing the first shuffled feature map to obtain desensitized image data.

[0097] In some embodiments, the first shuffling process may involve randomizing the order of the selected target frequency component channel feature maps. Randomization refers to shuffling the arrangement order of multiple feature maps. For example, the selected target frequency component channel feature maps are... Figure 6 The order of 61, 62, and 63 after randomization could be 63, 61, and 62.

[0098] In some embodiments, the normalization parameter refers to the parameter used when normalizing multiple target frequency component channel feature maps. During normalization, the normalization coefficients can be determined based on each frequency component channel feature map, ensuring that the normalization parameter used for each frequency component channel feature map is only related to that specific frequency component channel feature map and not to other frequency component feature maps. This increases the difficulty of inferring from the speech data. For example, suppose a frequency component channel feature map is inferred, but because the normalization parameters for each frequency component channel feature map are different, it is impossible to use the normalization parameters of the inferred frequency component channel map to infer other frequency component channel feature maps. This normalization process can also be called self-normalization.

[0099] In some embodiments, the normalization parameter can be the mean or variance of all values ​​in the frequency component channel feature map, or it can be the maximum or minimum value among all values ​​in the frequency component channel feature map. Normalization can be performed by removing the numerical values ​​of each element in the frequency component channel feature map using the normalization parameter, replacing the original values ​​with the quotient, thereby obtaining the normalized frequency component channel feature map.

[0100] By using the above-mentioned first shuffling process and corresponding normalization process, the original data of the selected target frequency component channel feature map can be made unavailable, thereby protecting the data privacy and security of the selected target frequency component channel feature map.

[0101] At 750, channel mixing is performed on the normalized first shuffle feature map.

[0102] Hybrid processing refers to performing calculations on two or more feature maps from multiple frequency component channel feature maps according to a preset calculation method. For example, the values ​​of corresponding elements in two or more frequency component channel feature maps can be calculated, and the calculated values ​​can be used as the values ​​of corresponding elements in the hybrid frequency component channel feature map. This allows two or more frequency component channel feature maps to be blended into a single frequency component channel feature map. The preset calculation method can be averaging, summing, or differencing.

[0103] In some embodiments, channel mixing can involve blending two adjacent frequency component channel feature maps. It should be noted that when combining feature maps, the combination rule should be the same for different frequency component channel feature maps. For example, starting with the first frequency component channel feature map, the current frequency component channel feature map is combined with its next adjacent frequency component channel feature map; that is, the first frequency component channel feature map is combined with the second frequency component channel feature map, and the second frequency component channel feature map is combined with the third frequency component channel feature map. Following this method, for M frequency component channel feature maps, M-1 frequency component channel feature maps can be obtained, thereby achieving dimensionality reduction.

[0104] In some embodiments, when mixing the feature maps of two adjacent frequency component channels, the number of selected target frequency component channel feature maps can be set to the feature dimension of the first feature data plus one. This processing method ensures that the feature dimension of the resulting desensitized image data is the same as the feature dimension of the first feature data, thus eliminating the need to modify the model architecture of the speech intent understanding model.

[0105] After channel mixing, the values ​​of each element in the frequency component channel feature map after channel mixing will change compared to the values ​​in the frequency component channel feature map before channel mixing. This can disrupt the relative magnitude of the values ​​of each element in the original frequency component channel feature map, thereby further increasing the difficulty of deducing the original speech data or its voiceprint information from the frequency component channel feature map.

[0106] After the channel mixing described above, at 760, the first shuffle feature map after channel mixing is subjected to a second shuffle process to obtain a second shuffle feature map; and the second shuffle feature map is then normalized. The second shuffle process and its normalization process are described in section 750 and will not be repeated here. In some embodiments, the data after normalizing the second shuffle feature map can be directly used as desensitized image data. In some embodiments, further processing is required on the data after normalizing the second shuffle feature map to obtain desensitized image data.

[0107] It should be noted that, Figure 3 and Figure 7 In the example, "submap data" is a frequency component channel feature map.

[0108] By utilizing the aforementioned second shuffling process and corresponding normalization, the original data of the channel feature map of the mixed frequency components can be made unavailable, thereby protecting the data privacy and security of the channel feature map of the mixed frequency components. It should be noted that in some embodiments, the first shuffling process can employ pseudo-random shuffling, and the second shuffling process can employ completely random shuffling.

[0109] exist Figure 7The data anonymization process shown, involving two shuffling and normalization processes, significantly increases the difficulty of brute-force attacks. For example, by performing a local cosine transform on the first feature data in 8x8 local image blocks, 64 frequency component channel feature maps can be constructed. After shuffling these frequency component channel feature maps, the placement of each small frequency component block (one frequency component channel feature map corresponds to one frequency component) in the transformation result is randomized, resulting in a randomized brute-force attack space of 64!, where "!" represents a factorial operation. Even if channel filtering is performed on the frequency component channel feature maps based on channel importance to determine, for example, 36 main feature maps, the brute-force attack space is still 36!. The anonymization process employs two randomization processes, making the brute-force attack space 36! * 36!, a value greater than the key cracking space of the 256-bit AES encryption algorithm. This makes it extremely difficult to deduce the original voice data through brute-force attacks. Furthermore, the normalization parameter depends solely on the corresponding frequency component channel feature map, and different frequency component channel feature maps have different normalization parameters. This makes it even more difficult to deduce the normalization parameter for each frequency component channel feature map. Additionally, channel mixing is performed on the frequency component channel feature maps, retaining only the mixed result. This disrupts the relative relationships between the values ​​of the frequency component channel feature maps, further increasing the difficulty of data decryption and thus enhancing the privacy and security protection of the voice data.

[0110] Back Figure 2 After obtaining the desensitized image data as described above, in step 220, the sub-image data in the obtained desensitized image data undergoes data augmentation-based image mixing processing to obtain the data-enhanced desensitized image data. In some embodiments, examples of data augmentation may include, but are not limited to, Mixup data augmentation, Instahide data augmentation, etc.

[0111] Mixup data augmentation typically involves two hyperparameters. One hyperparameter is the maximum weight coefficient W of the mixed images. max (The sum of all weight coefficients in the blended image needs to be 1). Typically, the maximum weight coefficient W... max The default value is 0.65. Preferably, the maximum weighting coefficient W is... max It can be set to 0.55 so that different images can contribute more data to the mixing, thus resulting in higher privacy protection. Another hyperparameter is the number of images, k, participating in the mixing operation. The larger the value of k, the more information is mixed, the stronger the privacy protection, but the lower the recognition rate. Optionally, the value of k is preferably 3.

[0112] Figure 8 An example flowchart of an image blending process 800 based on Mixup data enhancement according to an embodiment of this specification is shown.

[0113] like Figure 8 As shown in Figure 810, the sub-image data set in the obtained desensitized image data is shuffled k-1 times to obtain k sub-image data sets.

[0114] In step 820, based on the obtained k subgraph data sets, an image hypermatrix of size m*k is constructed. The first column of the constructed image hypermatrix corresponds to the original subgraph data set, and the remaining columns correspond to the subgraph data sets obtained after each shuffling process. m is the number of subgraph data sets in the original subgraph data set.

[0115] In 830, weight coefficients are randomly generated for each image in the image supermatrix.

[0116] In step 840, the weight coefficients of the images in the image hypermatrix are row-normalized so that the sum of the weight coefficients of each row of images is 1, and the weight coefficient of each image is no greater than W. max That is, after normalization, the maximum weight coefficient of each row of images cannot exceed W. max For example, it cannot exceed 0.55.

[0117] At 850, the images in each row of the image supermatrix are weighted and summed to obtain a hybrid image supermatrix of size m*1. The images in the obtained hybrid image supermatrix are desensitized image data after data augmentation.

[0118] Optionally, after obtaining the data-enhanced desensitized image data as described above, at step 230, each sub-image data in the data-enhanced desensitized image data is subjected to image size alignment processing so that the size of each sub-image data after image size alignment processing is the same as the size of the original image data.

[0119] After data anonymization and data augmentation, multiple sub-image data (e.g., frequency component channel feature maps) are obtained. The size of these sub-image data is usually different from the size of the input image data of the original image recognition network (image recognition model). To adapt to the input image data size of the original image recognition network, graph size alignment is required for the anonymized image data after data augmentation. Graph size alignment refers to upsampling or downsampling the size of each sub-image data to be the same as the input size of the original image recognition network. After graph size alignment, the only change to the existing image recognition network is to change the number of input channels to the number of channels corresponding to the anonymized image data, and then proceed with subsequent processing as a normal image recognition network, such as performing subsequent image recognition model training and image recognition processes.

[0120] As referred above Figures 1 to 8An image data processing method according to embodiments of this specification is described. The de-identified image data obtained using the above method can be provided to an image recognition model to predict the feature vector representation of the image data, or provided to an image recognition model for model training. The image recognition model can be any suitable machine learning model.

[0121] Figure 9 An example structural diagram of an image recognition model 900 according to an embodiment of this specification is shown. Figure 9 As shown, the image recognition model 900 may include an input layer 910, a feature extraction layer 920, and an output layer 930.

[0122] Input layer 910 can be used to receive signals referenced above. Figure 2 The image data described is processed to obtain desensitized image data after data enhancement.

[0123] In some embodiments, the input layer 910 may have multiple input channels, the number of which may be the same as the number of sub-map data (e.g., frequency component channel feature maps) in the desensitized image data, with each channel corresponding to one sub-map data.

[0124] In some embodiments, the number of input channels of the initially created image recognition model can be adjusted to match the number of sub-image data obtained by the image processing method described above.

[0125] In some embodiments, the number of sub-image data in the obtained desensitized image data can be set to match the number of channels in the original image recognition model by setting the number of target sub-image data after filtering. This allows the model architecture of the original image recognition model to be used without any adjustments.

[0126] The feature extraction layer 920 can be used to process the input desensitized image data to obtain the feature vector representation (or prediction vector) of the desensitized image data.

[0127] In some embodiments, the feature extraction layer can be a deep neural network, such as a CNN or RNN. The feature extraction layer can process each feature map (e.g., convolution, pooling) to obtain a more abstract feature vector representation.

[0128] The output layer 930 can convert the feature vector representation into the identification result of the target object corresponding to the desensitized image data.

[0129] The target object can refer to a living organism, an object, or a part thereof in the image. For example, a person, a face, an animal, or a building.

[0130] The identity recognition result can refer to the corresponding identity of the target object in the image, such as the identity of a person, the category of an animal, the name of a building, etc.

[0131] The output layer can transform the feature vector representation of the desensitized image data to obtain a predicted value, which can indicate the identity information of the task in the image, that is, the identity recognition result of the target object.

[0132] In some embodiments, the output layer may be a multilayer perceptron, a fully connected layer, etc., but this embodiment does not limit this.

[0133] Figure 10 An example flowchart of an image recognition model training method 1000 according to an embodiment of this specification is shown.

[0134] like Figure 10 As shown, at step 1010, image data (hereinafter referred to as training sample image data) and its labels are acquired to serve as training samples. The training sample image data may include face image data, and its labels may be identity information corresponding to the faces in the images, such as names. The labels for the training sample image data may be added manually or in other ways; this embodiment does not limit this. In some embodiments, the training sample image data and its labels may be acquired by reading from a database or calling a data interface.

[0135] In 1020, use the above reference. Figures 1 to 8 The image processing method described above processes image data to obtain desensitized image data after data augmentation. The desensitized image data after data augmentation includes a set of sub-image data, each of which has a weight coefficient.

[0136] In step 1030, the weight coefficients of each sub-image data of each desensitized image data are used to perform label mixing processing on the corresponding labels of each sub-image data of each desensitized image data.

[0137] In some embodiments, when performing label blending, if the sub-image data involved in the image blending process comes from the same category, the labels of each sub-image data involved in the image blending process remain unchanged. If at least some of the sub-image data involved in the image blending process comes from different categories, one non-zero value in the label of each sub-image data is adjusted to k non-zero values, each non-zero value corresponding to a weight coefficient of a sub-image data.

[0138] In 1040, the image recognition model was trained using the individual sub-image data of the desensitized image data and their labels after label mixing.

[0139] The training process of the image recognition model is a cyclical process. In each training round, a batch of training sample images of size `batch` is selected from the training sample image data. In the image mixing process described above, the mixed images used come from the same batch of training sample images. When the number of images participating in each mixing is set to `k`, the training sample images in the original batch are randomly shuffled by `k-1`, resulting in `k` batches of image data. Based on these `k` batches of image data, an image hypermatrix of size `batch*k` is constructed, where each item in the hypermatrix corresponds to one image, and the first column corresponds to the image data of the original batch. Next, weight coefficients of size `batch*k` are randomly generated for the image hypermatrix, and each weight coefficient is assigned to an item. Then, row normalization is performed so that the sum of the coefficients in each row is 1, and the maximum weight coefficient of the image in that row cannot exceed a specified value W. max Then, the weight coefficients of each row are used to perform a weighted summation on each row of the image supermatrix, finally obtaining a batch*1 size mixed image supermatrix, thus completing the image blending process.

[0140] In each training round, the image recognition model predicts the feature vector representation of the training sample images based on the anonymized image data, and determines the recognition result based on the predicted feature vector representation. For example, when the image data is face image data, the feature similarity between the predicted feature vector representation and each reference feature vector representation can be determined, and the recognition result, i.e., the identity information corresponding to the face, can be determined based on the determined feature similarity. Here, each reference feature vector representation corresponds to a specific identity information. When the feature similarity between the predicted feature vector representation and a reference feature vector representation is greater than a predetermined threshold, the face image is identified as having the identity information corresponding to that reference feature vector representation.

[0141] Furthermore, during each training round, a corresponding loss function can be constructed based on the recognition results output by the image recognition model and the corresponding labels. The model parameters of the image recognition model can then be adjusted according to the value of the loss function to reduce the difference between the recognition results and the labels. This training process is repeated until the model training termination condition is met, such as reaching the required number of training rounds or the difference between the recognition results and the labels meeting a predetermined requirement.

[0142] Optionally, in some embodiments, before using the desensitized image data and its labels after label mixing to train the image recognition model, the sub-image data in the desensitized image data can be subjected to graph size alignment processing so that the size of each sub-image data after graph size alignment processing is the same as the size of the original image data.

[0143] Figure 11An example flowchart of an image recognition method 1100 according to an embodiment of this specification is shown.

[0144] like Figure 11 As shown, at step 1110, the image data to be identified is acquired. The image data to be identified refers to the image from which the target object needs to be identified. Examples include face images and fingerprint images.

[0145] In some embodiments, the target object may include, for example, a face corresponding to specific identity information. Examples include, a face corresponding to the identity information of the payment account during facial recognition payment, a face corresponding to the identity information with unlocking permissions during facial recognition unlocking, and a face corresponding to the identity information with login permissions when logging into a terminal account.

[0146] In some embodiments, the processing device may acquire the image to be recognized through the camera of the terminal device, or by reading from a storage device, a database, or by calling a data interface.

[0147] At 1120, the image data to be identified is processed using the image processing method described above to obtain data-enhanced desensitized image data of the image data to be identified.

[0148] In 1130, the obtained desensitized image data is provided to the image recognition model to predict the feature vector representation of the image data to be recognized.

[0149] In step 1140, based on the predicted feature vector representation and the feature vector representation of the target object, it is determined whether the image data to be identified contains the target object. For example, the feature similarity between the predicted feature vector representation and the feature vector representation of the target object can be determined, and the recognition result can be determined based on the determined feature similarity. For example, if the feature similarity between the predicted feature vector representation and the feature vector representation of the target object is greater than a predetermined threshold, then the face image is considered to contain the target object, and thus the face image is identified as having the identity information corresponding to the target object. If the feature similarity between the predicted feature vector representation and the feature vector representation of the target object is not greater than the predetermined threshold, then the face image is considered not to contain the target object, and thus the face image is identified as not having the identity information corresponding to the target object.

[0150] In some embodiments, the feature vector representation of the target object can be pre-stored in a storage device. In some embodiments, the face image of the target object can be pre-processed. Figure 2 The image data processing obtains desensitized image data after data enhancement, and then the desensitized image data is processed using an image recognition model. The predicted feature vector output by the model's feature extraction layer is stored in a storage device as the feature vector representation of the target object.

[0151] In some embodiments, the feature vector representation of the target object can be obtained by reading from a database, storage device, or calling a data interface.

[0152] In some embodiments, when comparing the predicted feature vector representation with the feature vector representation of the target object, the distance between the two feature vector representations can be calculated, such as cosine distance, Euclidean distance, etc. The calculated distance between the two feature vector representations can represent the similarity between the two feature vector representations. If the similarity is greater than a preset value (e.g., 99%, 99.5%, etc.), it can be determined that the image to be identified contains the target object. If the similarity is not greater than the preset value, it can be determined that the image to be identified does not contain the target object.

[0153] In some embodiments, the above steps can be performed in a trusted execution environment to ensure that the entire process from image acquisition to image recognition is in a secure and trusted environment, thereby preventing the leakage of image and desensitized image data.

[0154] It should be noted that, Figure 11 The image recognition method 1100 shown can be executed by a terminal device or a server. Furthermore, the image recognition model can be deployed on a terminal device or a server.

[0155] It should be noted that the descriptions of the above processes are merely illustrative and do not limit the scope of application of the embodiments in this specification. Those skilled in the art can make various modifications and changes to the processes under the guidance of this specification. However, these modifications and changes are still within the scope defined by the embodiments in this specification. For example, changes to the process steps in the embodiments of this specification, such as adding preprocessing steps and storage steps.

[0156] Figure 12 An example block diagram of an image data processing apparatus 1200 according to an embodiment of this specification is shown. Figure 12 As shown, the image data processing device 1200 includes a data desensitization processing unit 1210 and an image mixing processing unit 1220.

[0157] The data desensitization processing unit 1210 is configured to perform frequency domain transformation-based data desensitization processing on image data to obtain desensitized image data. The obtained desensitized image data includes a set of sub-image data, where each sub-image data corresponds to a different frequency. The operation of the data desensitization processing unit 1210 can be referred to the above description. Figure 2 The operation described in 210.

[0158] The image blending processing unit 1220 is configured to perform data-enhanced image blending processing on the sub-image data in the obtained desensitized image data to obtain data-enhanced desensitized image data. The operation of the image blending processing unit 1220 can be referred to the above description. Figure 2 The operation described in 220.

[0159] Optionally, the image data processing apparatus 1200 may further include an image size alignment unit 1230. The image size alignment unit 1230 is configured to perform image size alignment processing on each sub-image data in the data-enhanced desensitized image data, such that the size of each sub-image data after image size alignment processing is the same as the size of the original image data. The operation of the image size alignment unit 1230 can be referred to the above description. Figure 2 The operation described in 230.

[0160] Figure 13 An example block diagram of a data desensitization processing unit 1300 according to an embodiment of this specification is shown. Figure 13 As shown, the data desensitization processing unit 1300 includes a local frequency domain transformation module 1310, a channel feature map construction module 1320, and a feature map filtering module 1330.

[0161] The local frequency domain transformation module 1310 is configured to perform local frequency domain transformation processing on the image data to obtain at least one feature map. Each feature map includes multiple elements and corresponds to a data block in the image data, with each element corresponding to a frequency in the frequency domain. The operation of the local frequency domain transformation module 1310 can be referred to the above reference. Figure 3 The operation described in 310.

[0162] The channel feature map construction module 1320 is configured to construct frequency component channel feature maps corresponding to each frequency using elements corresponding to each frequency in at least one feature map. The operation of the channel feature map construction module 1320 can be referenced above. Figure 3 The operation described in 320.

[0163] The feature map filtering module 1330 is configured to filter at least one target frequency component channel feature map from the constructed frequency component channel feature maps to obtain desensitized image data of the image data. The filtered target frequency component channel feature map is a key channel feature for image recognition. The operation of the feature map filtering module 1330 can be referred to the above. Figure 3 The operation described in 330.

[0164] Figure 14 An example block diagram of a data desensitization processing unit 1400 according to another embodiment of this specification is shown. Figure 14As shown, the data desensitization processing unit 1400 includes a local frequency domain transformation module 1410, a channel feature map construction module 1420, a feature map filtering module 1430, a first shuffling module 1440, a first normalization processing module 1450, a channel mixing processing module 1460, a second shuffling module 1470, and a second normalization processing module 1480.

[0165] The local frequency domain transformation module 1410 is configured to perform local frequency domain transformation processing on the image data to obtain at least one feature map. Each feature map includes multiple elements and corresponds to a data block in the image data, with each element corresponding to a frequency in the frequency domain. The operation of the local frequency domain transformation module 1410 can be referred to the above reference. Figure 7 The operation described in 710.

[0166] The channel feature map construction module 1420 is configured to construct frequency component channel feature maps for each frequency using elements corresponding to each frequency in at least one feature map. The operation of the channel feature map construction module 1420 can be referenced above. Figure 7 The operation described in 720.

[0167] The feature map filtering module 1430 is configured to filter at least one target frequency component channel feature map from the constructed frequency component channel feature map. The filtered target frequency component channel feature map is a key channel feature for image recognition. In some embodiments, the feature map filtering module 1430 can filter at least one target frequency component channel feature map from the constructed frequency component channel feature map based on channel importance or based on preset filtering rules. The operation of the feature map filtering module 1430 can be referred to the above reference. Figure 7 The operation described in 730.

[0168] The first shuffling module 1440 is configured to perform a first shuffling process on the target frequency component channel feature map to obtain a first shuffling feature map. The first normalization processing module 1450 is configured to normalize the first shuffling feature map. The operation of the first shuffling module 1440 and the first normalization processing module 1450 can be referred to the above. Figure 7 The operation described in 740.

[0169] The channel blending module 1460 is configured to perform channel blending processing on the normalized first shuffle feature map. The operation of the channel blending module 1460 can be referenced above. Figure 7 The operation described in 750.

[0170] The second shuffling module 1470 is configured to perform a second shuffling process on the first shuffling feature map after channel mixing processing to obtain a second shuffling feature map. The second normalization processing module 1480 is configured to normalize the second shuffling feature map. The operation of the first shuffling module 1470 and the second normalization processing module 1480 can be referred to the above. Figure 7 The operation described in 760.

[0171] Figure 15 An example block diagram of an image blending processing unit 1500 according to an embodiment of this specification is shown. Figure 15 As shown, the image mixing processing unit 1500 includes an image scrambling processing module 1510, an image supermatrix construction module 1520, a weight coefficient generation module 1530, a weight coefficient normalization module 1540, and an image mixing processing module 1550.

[0172] The image scrambling module 1510 is configured to scramble the sub-image data set of the desensitized image data k-1 times to obtain k sub-image data sets. The operation of the image scrambling module 1510 can be referenced above. Figure 8 The operation described in 810.

[0173] The image hypermatrix construction module 1520 is configured to construct an image hypermatrix of size m*k based on the obtained k sub-image data sets. The first column of the constructed image hypermatrix corresponds to the original sub-image data set, and m is the number of sub-image data in the original sub-image data set. The operation of the image hypermatrix construction module 1520 can be referenced above. Figure 8 The operation described in 820.

[0174] The weight coefficient generation module 1530 is configured to randomly generate weight coefficients for each image in the image supermatrix. The operation of the weight coefficient generation module 1530 can be referenced above. Figure 8 The operation described in 830.

[0175] The weight coefficient normalization module 1540 is configured to perform row normalization on the weight coefficients of the images in the image supermatrix, such that the sum of the weight coefficients of each row of images is 1, and the maximum coefficient of each image is no greater than W. max The operation of the weight coefficient normalization module 1540 can be referenced above. Figure 8 The operation described in 840.

[0176] The image blending processing module 1550 is configured to perform a weighted summation of each row of the image supermatrix to obtain a blended image supermatrix of size m*1. The images in the obtained blended image supermatrix are desensitized image data after data augmentation. The operation of the image blending processing module 1550 can be referred to the above. Figure 8 The operation described in 850.

[0177] Figure 16 An example block diagram of an image recognition model training apparatus 1600 according to an embodiment of this specification is shown. Figure 16 As shown, the image recognition model training device 1600 includes a training data acquisition unit 1610, an image data processing unit 1620, and a model training unit 1640.

[0178] The training data acquisition unit 1610 is configured to acquire image data and their labels to serve as training samples. The operation of the training data acquisition unit 1610 can be referenced above. Figure 10 The operation described in 1010.

[0179] The image processing unit 1620 is configured to use, for example Figure 2 The image data method described above processes image data to obtain data-enhanced de-identified image data of a face image. The data-enhanced de-identified image data includes a set of sub-image data, each sub-image data having a weight coefficient. The weight coefficients of each sub-image data in the de-identified image data are used to perform label mixing processing on the corresponding labels of each sub-image data in each de-identified image data. The operation of the image processing unit 1620 can be referred to [reference needed]. Figure 10 The operations described in 1020 and 1030.

[0180] The model training unit 1640 is configured to train an image recognition model using individual sub-image data of the desensitized image data and their labels after label mixing. The operation of the model training unit 1640 can be referenced above. Figure 10 The operation described in 1040.

[0181] Optionally, the image recognition model training apparatus 1600 may further include an image size alignment unit 1630. The image size alignment unit 1630 is configured to perform image size alignment processing on each sub-image data in the data-enhanced desensitized image data, such that the size of each sub-image data after image size alignment is the same as the size of the original image data. Accordingly, the model training unit 1640 is configured to train an image recognition model using the image size aligned desensitized image data and its labels after label mixing processing.

[0182] Figure 17An example block diagram of an image recognition device 1700 according to an embodiment of this specification is shown. Figure 17 As shown, the image recognition device 1700 includes an image data acquisition unit 1710, an image data processing unit 1720, a feature vector representation and prediction unit 1740, and an image recognition unit 1750.

[0183] The image data acquisition unit 1710 is configured to acquire image data to be recognized. The operation of the image data acquisition unit 1710 can be referenced above. Figure 11 The operation described in 1110.

[0184] The image data processing unit 1720 is configured to process image data using the image data method described above to obtain data-enhanced, desensitized image data. The operation of the image data processing unit 1720 can be referred to the above description. Figure 11 The operation described in 1120.

[0185] The feature vector representation prediction unit 1740 is configured to provide desensitized image data to an image recognition model to predict the feature vector representation of the image data. The operation of the feature vector representation prediction unit 1740 can be referenced above. Figure 11 The operation described in 1130.

[0186] The image recognition unit 1750 is configured to determine whether the image data to be recognized contains a target object based on the predicted feature vector representation and the feature vector representation of the target object. The operation of the image recognition unit 1750 can be referenced above. Figure 11 The operation described in 1140.

[0187] Optionally, the image recognition device 1700 may further include an image size alignment unit 1730. The image size alignment unit 1730 is configured to perform image size alignment processing on each sub-image data in the data-enhanced desensitized image data, such that the size of each sub-image data after image size alignment is the same as the size of the original image data. Accordingly, the feature vector representation prediction unit 1740 is configured to provide the image-size aligned desensitized image data to the image recognition model to predict the feature vector representation of the image data.

[0188] As per the above reference Figures 1 to 17 This specification describes an image data processing method, an image data processing apparatus, an image recognition model training method, an image recognition model training apparatus, an image recognition method, and an image recognition apparatus according to embodiments thereof. The image data processing apparatus, image recognition model training apparatus, and image recognition apparatus described above can be implemented in hardware, or in software, or a combination of hardware and software.

[0189] Figure 18 An example schematic diagram of an image data processing apparatus 1800 implemented on a computer system according to an embodiment of this specification is shown. Figure 18 As shown, the image data processing apparatus 1800 may include at least one processor 1810, a memory (e.g., non-volatile memory) 1820, a main memory 1830, and a communication interface 1840, and the at least one processor 1810, memory 1820, main memory 1830, and communication interface 1840 are connected together via a bus 1860. At least one processor 1810 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0190] In one embodiment, computer-executable instructions are stored in a memory that, when executed, cause at least one processor 1810 to: perform a frequency-domain transform-based data desensitization process on image data to obtain desensitized image data, the obtained desensitized image data including a set of sub-image data, each sub-image data in the sub-image data set corresponding to a different frequency; and perform a data-enhanced image mixing process on the sub-image data in the obtained desensitized image data to obtain data-enhanced desensitized image data.

[0191] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 1810 to perform the above-described combinations in the various embodiments of this specification. Figures 1-8 as well as Figures 12-15 The description includes various operations and functions.

[0192] Figure 19 An example schematic diagram of an image recognition model training apparatus 1900 implemented using a computer system according to an embodiment of this specification is shown. Figure 19 As shown, the image recognition model training apparatus 1900 may include at least one processor 1910, a memory (e.g., non-volatile memory) 1920, a main memory 1930, and a communication interface 1940, and the at least one processor 1910, memory 1920, main memory 1930, and communication interface 1940 are connected together via a bus 1960. At least one processor 1910 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0193] In one embodiment, computer-executable instructions are stored in memory that, when executed, cause at least one processor 1910 to: acquire image data and its labels serving as training samples; process the image data using the method described above to obtain data-enhanced desensitized image data of the image data; perform label mixing processing on the corresponding labels of the respective desensitized image data using weight coefficients of the respective desensitized image data; and train an image recognition model using the desensitized image data and its label-mixed labels.

[0194] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 1910 to perform the above-described combinations in the various embodiments of this specification. Figures 9-10 as well as Figure 16 The description includes various operations and functions.

[0195] Figure 20 An example schematic diagram of an image recognition device 2000 implemented on a computer system according to an embodiment of this specification is shown. Figure 20 As shown, the image recognition device 2000 may include at least one processor 2010, a memory (e.g., non-volatile memory) 2020, a RAM 2030, and a communication interface 2040, and the at least one processor 2010, memory 2020, RAM 2030, and communication interface 2040 are connected together via a bus 2060. The at least one processor 2010 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0196] In one embodiment, computer-executable instructions are stored in memory that, when executed, cause at least one processor 2010 to: acquire image data to be identified; process the image data using the method described above to obtain data-enhanced de-identified image data of the image data; provide the de-identified image data to an image recognition model to predict a predicted feature vector representation of the image data; and determine, based on the predicted feature vector representation and the feature vector representation of the target object, whether the image data to be identified contains the target object.

[0197] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 2010 to perform the above-described combinations in the various embodiments of this specification. Figure 11 as well as Figure 17 The description includes various operations and functions.

[0198] According to one embodiment, a program product, such as a machine-readable medium (e.g., a non-transitory machine-readable medium), is provided. The machine-readable medium may have instructions (i.e., the elements implemented in software as described above), which, when executed by a machine, cause the machine to perform the above-described combinations of the various embodiments of this specification. Figures 1-17 The various operations and functions described. Specifically, a system or apparatus equipped with a readable storage medium storing software program code that implements the functions of any of the embodiments described above, and enabling the computer or processor of the system or apparatus to read and execute the instructions stored in the readable storage medium.

[0199] In this case, the program code itself, which can be read from a readable medium, can perform the functions of any of the above embodiments. Therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0200] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0201] According to one embodiment, a computer program product is provided, the computer program product including a computer program, which, when executed by a processor, causes the processor to perform the above-described combinations of the various embodiments of this specification. Figures 1-17 The description includes various operations and functions.

[0202] Those skilled in the art will understand that the various embodiments disclosed above can be modified and varied without departing from the spirit of the invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

[0203] It should be noted that not all steps and units in the above process and system structure diagrams are mandatory; some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments can be a physical structure or a logical structure; that is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0204] In the above embodiments, the hardware units or modules can be implemented mechanically or electrically. For example, a hardware unit, module, or processor may include permanent dedicated circuitry or logic (such as a dedicated processor, FPGA, or ASIC) to perform the corresponding operation. The hardware unit or processor may also include programmable logic or circuitry (such as a general-purpose processor or other programmable processor), which can be temporarily configured by software to perform the corresponding operation. The specific implementation method (mechanical, dedicated permanent circuitry, or temporarily configured circuitry) can be determined based on cost and time considerations.

[0205] The specific embodiments described above with reference to the accompanying drawings are exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" compared to other embodiments. Specific details are included to provide an understanding of the described techniques. However, these techniques can be practiced without these specific details. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0206] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A training method for an image recognition model, comprising: Obtain image data and their labels to serve as training samples; The image data is subjected to data desensitization processing based on frequency domain transformation to obtain desensitized image data. The desensitized image data includes a set of sub-image data, each of which corresponds to a different frequency and has a weight coefficient. The sub-image data in the obtained desensitized image data are subjected to data augmentation-based image blending processing to obtain data augmented desensitized image data; Using the weight coefficients of each sub-image data in each desensitized image data, the corresponding labels of each sub-image data in each desensitized image data are mixed; as well as The image recognition model is trained using the sub-image data of the desensitized image data and the labels after label mixing.

2. The training method as described in claim 1, wherein, The data augmentation includes one of the following data augmentations: Mixup data enhancement; and Instahide data augmentation.

3. The training method as described in claim 2, wherein, The data augmentation includes Mixup data augmentation, where the number of images used for mixing is k, and the maximum weight coefficient of the mixed images is [value missing]. , The image blending process based on data augmentation for the sub-image data in the obtained desensitized image data includes: The sub-image data set of the desensitized image data is shuffled k-1 times to obtain k sub-image data sets; Based on the obtained k subgraph data sets, an image supermatrix of size m*k is constructed. The first column of the image supermatrix corresponds to the original subgraph data set, and m is the number of subgraph data in the original subgraph data set. Randomly generate weight coefficients for each image in the image hypermatrix; The weight coefficients of the images in the image hypermatrix are row-normalized so that the sum of the weight coefficients of each row of images is 1, and the weight coefficient of each image is no greater than 1. ;as well as The images in each row of the image supermatrix are weighted and summed to obtain a hybrid image supermatrix of size m*1. The images in the hybrid image supermatrix are desensitized image data after data augmentation.

4. The training method as described in claim 3, wherein, The The value of is 0.55, and the value of k is 3.

5. The training method as described in claim 1, further comprising: Each sub-image in the data-enhanced desensitized image data is subjected to image size alignment so that the size of each sub-image after image size alignment is the same as the size of the original image data.

6. The training method as described in claim 1, wherein, Data desensitization processing based on frequency domain transformation of image data includes: The image data is subjected to local frequency domain transformation to obtain at least one feature map. Each feature map includes multiple elements and corresponds to a data block in the image data. Each element corresponds to a frequency in the frequency domain. Using the elements corresponding to each frequency in the at least one feature map, a frequency component channel feature map corresponding to each frequency is constructed; and At least one target frequency component channel feature map is selected from the constructed frequency component channel feature map to obtain the desensitized image data of the image data. The selected target frequency component channel feature map is the key channel feature for image recognition.

7. The training method as described in claim 6, wherein, After selecting at least one target frequency component channel feature map from the constructed frequency component channel feature map, the method further includes: The target frequency component channel feature map is subjected to a first shuffle process to obtain a first shuffled feature map; and The first shuffle feature map is normalized to obtain the desensitized image data of the image data.

8. The training method as described in claim 6, wherein, Selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps includes: Based on channel importance or based on preset filtering rules, at least one target frequency component channel feature map is selected from the constructed frequency component channel feature map.

9. The training method as described in claim 7, wherein, After normalizing the first shuffle feature map, the method further includes: Channel blending is performed on the first shuffle feature map after normalization. The first shuffle feature map, after channel mixing, is subjected to a second shuffle process to obtain the second shuffle feature map; and The second shuffle feature map is normalized to obtain the desensitized image data of the image data.

10. The training method as described in claim 9, wherein, The number of target frequency component channel feature maps is equal to the feature dimension of the first feature data plus one. Channel mixing processing of the first shuffle feature map includes: Channel mixing is performed on two adjacent frequency component channel feature maps in the first shuffle feature map.

11. The training method as described in claim 1, wherein, When the sub-image data involved in image blending comes from the same category, the labels of each sub-image data involved in image blending remain unchanged. When at least some of the sub-image data involved in image blending comes from different categories, the non-zero values ​​in the labels of each sub-image data are adjusted to k non-zero values, and each non-zero value corresponds to a weight coefficient of a sub-image data.

12. The training method according to claim 1, wherein, The image data includes facial image data, as well as the label or recognition result indicating the identity information corresponding to the face; The image recognition model is used to predict the feature vector representation of the image data based on the desensitized image data of the image data, and to determine the recognition result based on the predicted feature vector representation.

13. A training device for an image recognition model, comprising: The training data acquisition unit acquires image data and their labels to serve as training samples. The image processing unit performs data desensitization processing on the image data based on frequency domain transformation to obtain desensitized image data. The desensitized image data includes a set of sub-image data, each sub-image data in the set corresponding to a different frequency and having a weight coefficient. The unit then performs image mixing processing based on data augmentation on the sub-image data in the obtained desensitized image data to obtain data-enhanced desensitized image data. Additionally, using the weight coefficients of each sub-image data in each desensitized image data set, the unit performs label mixing processing on the corresponding labels of each sub-image data in each desensitized image data set. as well as The model training unit uses the sub-image data of the desensitized image data and its labels after label mixing to train the image recognition model.

14. An image recognition method, comprising: Acquire the image data to be recognized; The image data is subjected to data desensitization processing based on frequency domain transformation to obtain desensitized image data, which includes a set of sub-image data, each of which corresponds to a different frequency. The sub-image data in the obtained desensitized image data are subjected to data augmentation-based image blending processing to obtain data augmented desensitized image data; The desensitized image data is provided to an image recognition model to predict the feature vector representation of the image data; as well as Based on the predicted feature vector representation and the feature vector representation of the target object, it is determined whether the image data to be identified contains the target object. The data enhancement includes Mixup data enhancement, where the number of images used for mixing is k, and the maximum weight coefficient of the mixed images is [value missing]. , The image blending process based on data augmentation for the sub-image data in the obtained desensitized image data includes: The sub-image data set of the desensitized image data is shuffled k-1 times to obtain k sub-image data sets; Based on the obtained k subgraph data sets, an image supermatrix of size m*k is constructed. The first column of the image supermatrix corresponds to the original subgraph data set, and m is the number of subgraph data in the original subgraph data set. Randomly generate weight coefficients for each image in the image hypermatrix; The weight coefficients of the images in the image hypermatrix are row-normalized so that the sum of the weight coefficients of each row of images is 1, and the weight coefficient of each image is no greater than 1. ;as well as The images in each row of the image supermatrix are weighted and summed to obtain a hybrid image supermatrix of size m*1. The images in the hybrid image supermatrix are desensitized image data after data augmentation.

15. The image recognition method as described in claim 14, further comprising: Each sub-image in the data-enhanced desensitized image data is subjected to image size alignment so that the size of each sub-image after image size alignment is the same as the size of the original image data.

16. The image recognition method as described in claim 14, wherein, Data desensitization processing based on frequency domain transformation of image data includes: The image data is subjected to local frequency domain transformation to obtain at least one feature map. Each feature map includes multiple elements and corresponds to a data block in the image data. Each element corresponds to a frequency in the frequency domain. Using the elements corresponding to each frequency in the at least one feature map, a frequency component channel feature map corresponding to each frequency is constructed; and At least one target frequency component channel feature map is selected from the constructed frequency component channel feature map to obtain the desensitized image data of the image data. The selected target frequency component channel feature map is the key channel feature for image recognition.

17. The image recognition method as described in claim 16, wherein, After selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps, the image recognition method further includes: The target frequency component channel feature map is subjected to a first shuffle process to obtain a first shuffled feature map; and The first shuffle feature map is normalized to obtain the desensitized image data of the image data.

18. The image recognition method as described in claim 16, wherein, Selecting at least one target frequency component channel feature map from the constructed frequency component channel feature maps includes: Based on channel importance or based on preset filtering rules, at least one target frequency component channel feature map is selected from the constructed frequency component channel feature map.

19. The image recognition method as described in claim 17, wherein, After normalizing the first shuffling feature map, the image recognition method further includes: Channel blending is performed on the first shuffle feature map after normalization. The first shuffle feature map, after channel mixing, is subjected to a second shuffle process to obtain the second shuffle feature map; and The second shuffle feature map is normalized to obtain the desensitized image data of the image data.

20. The image recognition method as described in claim 19, wherein, The number of target frequency component channel feature maps is equal to the feature dimension of the first feature data plus one. Channel mixing processing of the first shuffle feature map includes: Channel mixing is performed on two adjacent frequency component channel feature maps in the first shuffle feature map.

21. The image recognition method as described in claim 14, wherein, The image data to be identified includes facial image data, and the target object includes facial image data with specific identity information.

22. An image recognition device, comprising: Image data acquisition unit acquires image data to be recognized; The image data processing unit performs data desensitization processing based on frequency domain transformation on the image data to obtain desensitized image data. The desensitized image data includes a set of sub-image data, each sub-image data in the set of sub-image data corresponding to a different frequency. The unit then performs image mixing processing based on data enhancement on the sub-image data in the obtained desensitized image data to obtain data-enhanced desensitized image data. The feature vector representation prediction unit provides the desensitized image data to the image recognition model to predict the feature vector representation of the image data. as well as The image recognition unit determines whether the image data to be recognized contains the target object based on the predicted feature vector representation and the feature vector representation of the target object. The data enhancement includes Mixup data enhancement, where the number of images used for mixing is k, and the maximum weight coefficient of the mixed images is [value missing]. The image data processing unit is configured as follows: The sub-image data set of the desensitized image data is shuffled k-1 times to obtain k sub-image data sets; Based on the obtained k subgraph data sets, an image supermatrix of size m*k is constructed. The first column of the image supermatrix corresponds to the original subgraph data set, and m is the number of subgraph data in the original subgraph data set. Randomly generate weight coefficients for each image in the image hypermatrix; The weight coefficients of the images in the image hypermatrix are row-normalized so that the sum of the weight coefficients of each row of images is 1, and the weight coefficient of each image is no greater than 1. ;as well as The images in each row of the image supermatrix are weighted and summed to obtain a hybrid image supermatrix of size m*1. The images in the hybrid image supermatrix are desensitized image data after data augmentation.

23. A training device for an image recognition model, comprising: At least one processor, Memory coupled to the at least one processor, and A computer program stored in the memory, which is executed by the at least one processor to implement the training method for an image recognition model as described in any one of claims 1 to 12.

24. An image recognition device, comprising: At least one processor, Memory coupled to the at least one processor, and A computer program stored in the memory, which is executed by the at least one processor to implement the image recognition method as described in any one of claims 14 to 21.

25. A computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform a training method for an image recognition model as described in any one of claims 1 to 12, or to perform an image recognition method as described in any one of claims 14 to 21.

26. A computer program product comprising a computer program that is executed by a processor to implement a training method for an image recognition model as described in any one of claims 1 to 12, or to implement an image recognition method as described in any one of claims 14 to 21.

Images