Image data desensitization method, device and equipment

Abstract

The invention relates to the technical field of image processing, in particular to an image data desensitization method, device and equipment, and the method comprises the steps: extracting a to-be-processed image or a RAW domain image of a to-be-processed video stream; sequentially performing wide dynamic range processing, first Bayer noise processing and tone mapping processing on the RAW domain image to obtain a first processed image; obtaining a dynamic ROI detection model and a privacy grading model; based on the first processing image and the dynamic ROI detection model, a second processing image is output, and the second processing image is that a privacy area is marked on the first processing image; based on the second processing image and a privacy grading model, obtaining a privacy level comprehensive score of each privacy area; and determining a corresponding desensitization means based on the privacy level comprehensive score of each privacy area so as to carry out pixel level privacy processing on the second processing image, and realizing end-to-end security privacy protection by processing the RAW domain image of the video or the image.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method, device and equipment for desensitizing image data. Background Art

[0002] With the widespread use of video surveillance, smart security, and mobile devices, personal privacy information contained in image and video data, such as faces, license plates, and ID card information, requires special protection. Traditional privacy masking methods, such as blurring, mosaics, and color block overlays, typically rely on manual post-processing. This involves workers or algorithms detecting and masking sensitive information after the ISP (Image Signal Processor) outputs the RGB image or video encoding.

[0003] Moreover, traditional methods usually perform desensitization at the ISP output (RGB domain) or after video encoding, while the original image before this can still be stored or transmitted, which is prone to the risk of malicious recovery.

[0004] Therefore, how to use desensitization methods to maximize the protection of images from malicious restoration is a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method, apparatus and device for desensitizing image data that overcome the above problems or at least partially solve the above problems.

[0006] In a first aspect, the present invention provides a method for desensitizing image data, comprising: Obtaining a RAW domain image of an image to be processed or a video stream to be processed; performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image in sequence to obtain a first processed image; Obtain dynamic ROI detection model and privacy classification model; Outputting a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first processed image; Obtaining a comprehensive privacy score for each privacy area based on the second processed image and the privacy grading model; Based on the comprehensive privacy level score of each privacy area, the corresponding desensitization method is determined and pixel-level privacy processing is performed.

[0007] Preferably, after obtaining the RAW domain image of the image to be processed or the video stream to be processed, the method further includes: Perform ISP preprocessing on the RAW domain image to obtain a preprocessed image.

[0008] Preferably, after performing wide dynamic range processing on the RAW domain image, the method further includes: Mark motion blurred areas.

[0009] Preferably, obtaining a dynamic ROI detection model includes: Acquire a first historical RAW domain image marked with a historical privacy area and a second historical RAW domain image marked with a motion blur area; The first historical RAW domain image and the second historical RAW domain image are input into a neural network model for training to obtain a dynamic ROI detection model for marking privacy areas and motion blur areas.

[0010] Preferably, obtaining a dynamic ROI detection model further includes: The parameters of the dynamic ROI detection model are adjusted so that subsequent processing is skipped for motion blurred areas that are not discernible to the human eye.

[0011] Preferably, based on the second processed image and the privacy grading model, a comprehensive privacy level score of each privacy area is obtained, including: After performing wide dynamic range processing on the RAW domain image, the method further includes: Using a second Bayer noise process to obtain a third processed image, wherein precision of the second Bayer noise process is lower than precision of the first Bayer noise process; Mapping the privacy area marked in the second processed image to the third processed image to obtain a fourth processed image; The fourth processed image is input into a privacy grading model to obtain a comprehensive privacy level score for each privacy area.

[0012] Preferably, based on the comprehensive privacy level score of each privacy area, a corresponding desensitization method is determined to perform pixel-level privacy processing, including: Determining a corresponding privacy level based on the comprehensive privacy level scores of the privacy areas; Determine corresponding desensitization measures based on the privacy level; Based on corresponding desensitization measures, pixel-level privacy processing is performed on the second processed image.

[0013] Preferably, the desensitization means include: Mild blurring, pixelation, non-decipherable mosaic overlays, and decipherable mosaic overlays of private areas.

[0014] In a second aspect, the present invention further provides a device for desensitizing image data, comprising: A first acquisition module is used to acquire a RAW domain image of an image to be processed or a video stream to be processed; an image processing module, configured to sequentially perform wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image to obtain a first processed image; The second acquisition module is used to obtain a dynamic ROI detection model and a privacy classification model; a marking module, configured to output a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first processed image; an obtaining module, configured to obtain a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model; The desensitization processing module is used to determine the corresponding desensitization means based on the comprehensive privacy level score of each privacy area to perform pixel-level privacy processing.

[0015] In a third aspect, the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect when executing the program.

[0016] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the method described in the first aspect when the program is executed by a processor.

[0017] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages: The present invention provides a method for desensitizing image data, comprising: obtaining a RAW domain image of an image to be processed or a video stream to be processed; performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image in sequence to obtain a first processed image; obtaining a dynamic ROI detection model and a privacy grading model; outputting a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first image; obtaining a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model; determining a corresponding desensitization measure based on the comprehensive privacy level score for each privacy area, so as to perform pixel-level privacy processing on the second processed image, and achieving end-to-end security and privacy protection by processing the RAW domain image of a video or image. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. Throughout the drawings, the same reference figures denote the same components. In the drawings: Figure 1A schematic diagram showing the steps of a method for desensitizing image data according to an embodiment of the present invention is shown; Figure 2 A schematic structural diagram of an apparatus for desensitizing image data according to an embodiment of the present invention is shown; Figure 3 A schematic diagram of the structure of a computer device for implementing a method for desensitizing image data in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0020] Example 1: The embodiment of the present invention provides a method for desensitizing image data, such as Figure 1 As shown, including: S101, obtaining a RAW domain image of an image to be processed or a video stream to be processed; S102, sequentially performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image to obtain a first processed image; S103, obtaining a dynamic ROI detection model and a privacy classification model; S104, outputting a second processed image based on the first processed image and the dynamic ROI detection model, where the second processed image is the first processed image with the privacy area marked; S105, obtaining a comprehensive privacy score for each privacy area based on the second processed image and the privacy grading model; S106: Based on the comprehensive privacy level score of each privacy area, determine the corresponding desensitization method to perform pixel-level privacy processing.

[0021] The overall idea of ​​this solution is to extract RAW domain images from the image to be processed or the video stream to be processed, and then process them through wide dynamic range processing (WDR), first Bayer noise processing and tone mapping. Finally, it goes through a dynamic ROI detection model, a privacy classification model and pixel-level RAW domain desensitization processing, thereby achieving end-to-end privacy protection and preventing the original data from leaving the device.

[0022] In a specific embodiment, S101 obtains a RAW domain image of an image or video stream to be processed. Specifically, the RAW domain image of an image or video is the raw photoelectric signal captured by a sensor after preliminary processing and without further processing. It is the starting point of the data lifecycle and retains the original state at the time of data generation. It may contain noise, redundancy, errors, or unstructured information.

[0023] After the RAW domain image is obtained, executing S102, performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image in sequence to obtain a first processed image; Before this series of processing, it also includes: ISP preprocessing is performed on the RAW image to produce a preprocessed image. This ISP preprocessing includes black level correction (BLC), which eliminates sensor dark current noise and ensures signal linearity by calibrating the offset; lens shading correction (LSC), which compensates for edge brightness attenuation and color deviation caused by lens optical characteristics; and bad pixel correction (BPC), which detects and repairs bad or abnormal sensor pixels to avoid fixed noise points in the image. Of course, other processing steps are also included, which are not detailed here.

[0024] Next, execute S102 to perform wide dynamic range processing (WDR) on the pre-processed RAW domain image. Specifically, a wide dynamic range image is generated by synthesizing multiple frames of dynamic range images (LDR). This allows the capture of clear and realistic images under complex lighting conditions such as strong light sources, backlighting, and backlighting, without darkening or deteriorating the image due to shadows, glare, reflections, and sunlight.

[0025] Next, the first Bayer noise processing is performed. Since the RAW domain is mainly Poisson noise and Gaussian noise, the noise shape is clear and is not affected by the images of other back-end modules. By setting the RAW domain noise reduction intensity to the maximum, the noise reduction is cleaner, even if details are lost.

[0026] Next, tone mapping is performed. Specifically, lightweight tone mapping is used to adaptively resize the dynamic range of the noise-reduced RAW image to a more balanced and detectable distribution, resulting in the first processed image. Tone mapping compresses high dynamic range (HDR) images to a displayable range for low dynamic range (LDR) devices while preserving detail.

[0027] Next, execute S103 to obtain a dynamic ROI detection model and a privacy classification model.

[0028] The dynamic ROI detection model is specifically used to mark the privacy area and motion blur area of ​​the processed RAW domain image.

[0029] Specifically, by obtaining a first historical RAW domain image marked with a historical privacy area and a second historical RAW domain image marked with a motion blur area; the first historical RAW domain image and the second historical RAW domain image are input into a neural network model for training, and a dynamic ROI detection model for marking privacy areas and motion blur areas is obtained.

[0030] Among them, in order to resist the influence of motion blur on the accuracy of dynamic ROI detection, the motion blur marker (motion mask) is integrated. With the help of the image information of short-exposure bright areas and long-exposure dark areas in the multi-exposure frames output by wide dynamic range (WDR) processing, the motion blur marker (motion mask) is generated through optical flow estimation or traditional inter-frame difference method to mark the motion blurred area.

[0031] When training the dynamic ROI detection model, As an attention weight, the dynamic ROI detection model is forced to focus on the static low motion blur area and reduce the detection weight of the motion area. Therefore, the motion marker weighting is introduced into the loss function:

[0032] in, Motion perception loss. is the standard loss, is a weighting coefficient used to balance the weight between motion perception loss and standard loss, is the predicted value, is the true value, The value of is used to identify the motion intensity of different areas. Low motion intensity values ​​represent high-motion areas, such as fast-moving objects or blurred parts, and high motion intensity values ​​represent static or low-motion areas, such as clear and stable areas.

[0033] To mitigate motion blur, we added motion blur enhancement data to the training process of the dynamic ROI detection model, improving the model's scene adaptability at a minimal cost. We added motion blur samples that simulate camera shake and rapidly moving objects to improve the model's adaptability to motion-blurred images. We also used motion_mask information to annotate motion-blurred areas.

[0034] By adjusting the parameters of the dynamic ROI detection model, subsequent processing is skipped for motion blurred areas that are not discernible to the human eye to avoid invalid processing.

[0035] The dynamic ROI detection model then outputs the confidence level of the privacy regions in the first processed image. For example, privacy regions such as faces and license plates in the first processed image are weighted and scored based on their size and number, with the score being recorded as obj_score.

[0036] The privacy classification model is also obtained through model training and is pre-classified into 5 levels from low to high according to the privacy level of the image scene: Level 1 public open scenes, such as natural scenery with basically no people / cars.

[0037] L2 semi-public scenes, such as parks, shopping malls, etc.

[0038] L3 sensitive public scenarios, such as public transportation, schools, hospitals, etc.

[0039] L4 high-privacy scenarios, such as dressing rooms, homes, etc.

[0040] L5 confidential scenes, such as areas where confidential targets appear.

[0041] By using a large amount of sample data to train the model, a privacy grading model that can identify the privacy level of image scenes is obtained.

[0042] Next, S104 is executed to output a second processed image based on the first processed image and the dynamic ROI detection model. The second processed image is a first processed image with the privacy area marked.

[0043] Specifically, the first processed image is input into the dynamic ROI detection model, so that the privacy area in the first processed image can be identified and determined, that is, the second processed image with the privacy area marked, and the privacy area is marked with the confidence obj_score of each privacy area.

[0044] Next, step S105 is executed to obtain a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model.

[0045] Specifically, after performing wide dynamic range processing on the RAW domain image, the method further includes: performing a second Bayer noise processing to obtain a third processed image, wherein the second Bayer noise processing accuracy is lower than the first Bayer processing accuracy; Mapping the marked privacy area in the second processed image to the third processed image to obtain a fourth processed image; The fourth processed image is input into the privacy grading model to obtain a comprehensive privacy level score for each privacy area.

[0046] The processing accuracy of the second Bayer noise processing is lower than the processing accuracy of the first Bayer noise processing, so as to ensure that the third processed image retains more information for subsequent processing.

[0047] Since the second processed image has a privacy area marked, while the third processed image does not, but the third processed image contains more information that is not filtered out by the second Bayer noise processing, the marked privacy area in the second processed image is mapped to the third processed image to obtain a fourth processed image. The fourth processed image not only contains more information that is not filtered out by the second Bayer noise processing, but also contains the marked privacy area.

[0048] Next, the fourth processed image is input into the privacy grading model to obtain a comprehensive privacy level score for each privacy area.

[0049] Specifically, the 5-dimensional probability vector scene_probs=[P1, P2, P3, P4, P5] is output, and each type of scene privacy level is assigned a probability weight. The higher the privacy level, the greater the weight. The weighted score is recorded as scene_score.

[0050] By combining the obj_score output by the dynamic ROI detection model, the scene pre-classification is corrected. For example, if a scene is pre-classified as a public open scene but people or cars appear, it is corrected to a semi-public scene. The weighted privacy level combined_score is obtained, and finally the score is mapped to the scene privacy level L1 to L5.

[0051] combined_score=scene_score*alpha1+obj_score*alpha2 Among them, alpha1 and alpha2 are adjusted according to user wishes.

[0052] The higher the overall privacy level score, the greater the desensitization intensity.

[0053] Finally, execute S106 to determine the corresponding desensitization measures based on the comprehensive privacy level scores of each privacy area and perform pixel-level privacy processing.

[0054] Specifically, based on the comprehensive privacy level score of each privacy area, the corresponding privacy level is determined; Determine the appropriate desensitization measures based on the privacy level; Based on the corresponding desensitization means, pixel-level privacy processing is performed on the second processed image.

[0055] In a specific embodiment, a privacy level comprehensive score corresponding to the fourth processed image is obtained, and different desensitization processing methods are adopted according to the privacy level comprehensive score. The desensitization methods include: slight blurring, pixelation, non-decryptable mosaic coverage, and decryptable mosaic coverage of the privacy area. The following is an introduction to the desensitization methods corresponding to different privacy levels: First of all, these desensitization methods are all based on the characteristics of the Bayer array to avoid color distortion caused by demosaicing post-processing.

[0056] Specifically, level L1: no desensitization, suitable for public open scenes, uses original data (RAW domain image), retains the complete Bayer array, and outputs unprocessed original images.

[0057] Level L2: Mild desensitization, using a semi-public scene, using channel-by-channel Gaussian blur, using a 5×5 Gaussian kernel for each channel, retaining color consistency after blurring, irreversible restoration, and overall light blurring. The outline of the identified privacy area is visible, but the actual content is indistinguishable.

[0058] Level L3: Moderate desensitization, suitable for sensitive public scenes. It uses pixel ratio averaging and takes the average of the same channel in a 4×4 Bayer block. It performs irreversible restoration and pixelates the private area.

[0059] Level L4: Moderate desensitization, suitable for high-privacy scenarios, using dynamic mosaics and color block coverage. 2×2 Bayer block mosaics are used for private areas such as faces / license plates, with superimposed black rectangular blocks. This is irreversible restoration, with mosaics and color block coverage for private areas, and a blurred background.

[0060] Level L5: Complete shielding, suitable for confidential scenarios. AES encrypted metadata is used to mark confidential privacy areas. Confidential privacy areas, such as GPS coordinates or military restricted areas, can be restored to their original data using a key. Other areas are still covered with a 2×2 Bayer irreversible mosaic.

[0061] In this way, the image data is desensitized to avoid malicious restoration.

[0062] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages: The present invention provides a method for desensitizing image data, comprising: obtaining a RAW domain image of an image to be processed or a video stream to be processed; performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image in sequence to obtain a first processed image; obtaining a dynamic ROI detection model and a privacy grading model; outputting a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first image; obtaining a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model; determining a corresponding desensitization measure based on the comprehensive privacy level score for each privacy area, so as to perform pixel-level privacy processing on the second processed image, and achieving end-to-end security and privacy protection by processing the RAW domain image of a video or image.

[0063] Example 2: Based on the same inventive concept, an embodiment of the present invention also provides a device for desensitizing image data, such as Figure 2 Shown, including: A first acquisition module 201 is configured to acquire a RAW domain image of an image to be processed or a video stream to be processed; An image processing module 202 is configured to sequentially perform wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image to obtain a first processed image; A second acquisition module 203 is used to acquire a dynamic ROI detection model and a privacy classification model; The marking module 204 is configured to output a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first processed image; an obtaining module 205 for obtaining a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model; The desensitization processing module 206 is used to determine corresponding desensitization measures based on the comprehensive privacy level scores of the privacy areas to perform pixel-level privacy processing.

[0064] In an optional embodiment, the method further includes a pre-processing module configured to: After obtaining a RAW domain image of an image to be processed or a video stream to be processed, ISP preprocessing is performed on the RAW domain image to obtain a preprocessed image.

[0065] In an optional embodiment, the system further includes a marking module configured to: After performing wide dynamic range processing on the RAW domain image, the motion blur area is marked.

[0066] In an optional implementation, the second acquisition module 203 is configured to: Acquire a first historical RAW domain image marked with a historical privacy area and a second historical RAW domain image marked with a motion blur area; The first historical RAW domain image and the second historical RAW domain image are input into a neural network model for training to obtain a dynamic ROI detection model for marking privacy areas and motion blur areas.

[0067] In an optional implementation, the second acquisition module 203 is configured to: The parameters of the dynamic ROI detection model are adjusted so that subsequent processing is skipped for motion blurred areas that are not discernible to the human eye.

[0068] In an optional embodiment, module 205 is obtained for: After performing wide dynamic range processing on the RAW domain image, the method further includes: Using a second Bayer noise process to obtain a third processed image, wherein precision of the second Bayer noise process is lower than precision of the first Bayer noise process; Mapping the privacy area marked in the second processed image to the third processed image to obtain a fourth processed image; The fourth processed image is input into a privacy grading model to obtain a comprehensive privacy level score for each privacy area.

[0069] In an optional embodiment, the desensitization processing module 206 is configured to: Determining a corresponding privacy level based on the comprehensive privacy level scores of the privacy areas; Determine corresponding desensitization measures based on the privacy level; Based on corresponding desensitization measures, pixel-level privacy processing is performed on the second processed image.

[0070] In an optional embodiment, the desensitization means comprises: Mild blurring, pixelation, non-decipherable mosaic overlays, and decipherable mosaic overlays of private areas.

[0071] Example 3: Based on the same inventive concept, an embodiment of the present invention provides a computer device, such as Figure 3 As shown, it includes a memory 304, a processor 302, and a computer program stored in the memory 304 and executable on the processor 302. When the processor 302 executes the program, the steps of the above-mentioned method for desensitizing image data are implemented.

[0072] Among them, Figure 3 In the present invention, a bus architecture (represented by bus 300) is shown. Bus 300 may include any number of interconnected buses and bridges. Bus 300 links various circuits, including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also link various other circuits, such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and, therefore, will not be described further herein. Bus interface 306 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same component, namely a transceiver, which provides a means for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 may be used to store data used by processor 302 when performing operations.

[0073] Example 4: Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above-mentioned method for desensitizing image data when executed by a processor.

[0074] The algorithm and display provided herein are not inherently related to any particular computer, virtual system or other device. Various general-purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing this type of system. In addition, the present invention is not directed to any specific programming language. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the above description of specific languages ​​is for the purpose of disclosing the best mode of the present invention.

[0075] In the description provided herein, numerous specific details are described. However, it is understood that embodiments of the present invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0076] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting an intention that the claimed invention requires more features than those explicitly recited in each embodiment. Rather, as reflected in each embodiment, inventive aspects lie in fewer than all the features of the individual embodiments previously disclosed. Accordingly, the claims that follow the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of the present invention.

[0077] Those skilled in the art will appreciate that the modules in the devices of the embodiments can be adaptively modified and installed in one or more devices different from the embodiments. The modules, units, or components in the embodiments can be combined into a single module, unit, or component, and furthermore, they can be divided into multiple sub-modules, sub-units, or sub-components. All features disclosed in this specification (including the accompanying claims, abstract, and drawings), and all processes or units of any method or device disclosed therein, can be combined in any combination, unless at least some of such features and / or processes or units are mutually exclusive. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0078] Furthermore, those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of the present invention and to form different embodiments. For example, in a specific embodiment, any one of the claimed embodiments may be used in any combination.

[0079] The various component embodiments of the present invention may be implemented in hardware, as software modules running on one or more processors, or as a combination thereof. Those skilled in the art will appreciate that, in practice, a microprocessor or digital signal processor (DSP) may be used to implement some or all of the functions of some or all of the components of the apparatus for desensitizing image data or a computer device according to embodiments of the present invention. The present invention may also be implemented as an apparatus or device program (e.g., a computer program or computer program product) for performing part or all of the methods described herein. Such a program implementing the present invention may be stored on a computer-readable medium or in the form of one or more signals. Such signals may be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0080] It should be noted that the above embodiments illustrate rather than limit the invention, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising several different elements and by means of appropriately programmed computers. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

Claims

1. A method for desensitizing image data, characterized in that: include: Obtaining a RAW domain image of an image to be processed or a video stream to be processed; performing wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image in sequence to obtain a first processed image; Obtain dynamic ROI detection model and privacy classification model; Outputting a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first processed image; Obtaining a comprehensive privacy score for each privacy area based on the second processed image and the privacy grading model; Based on the comprehensive privacy level score of each privacy area, the corresponding desensitization method is determined and pixel-level privacy processing is performed.

2. The method according to claim 1, wherein After obtaining the RAW domain image of the image to be processed or the video stream to be processed, the following steps are also included: Perform ISP preprocessing on the RAW domain image to obtain a preprocessed image.

3. The method according to claim 1, wherein After performing wide dynamic range processing on the RAW domain image, the method further includes: Mark motion blurred areas.

4. The method according to claim 3, wherein Get the dynamic ROI detection model, including: Acquire a first historical RAW domain image marked with a historical privacy area and a second historical RAW domain image marked with a motion blur area; The first historical RAW domain image and the second historical RAW domain image are input into a neural network model for training to obtain a dynamic ROI detection model for marking privacy areas and motion blur areas.

5. The method according to claim 4, wherein Obtaining a dynamic ROI detection model also includes: The parameters of the dynamic ROI detection model are adjusted so that subsequent processing is skipped for motion blurred areas that are not discernible to the human eye.

6. The method according to claim 1, wherein Based on the second processed image and the privacy grading model, a comprehensive privacy level score of each privacy area is obtained, including: After performing wide dynamic range processing on the RAW domain image, the method further includes: Using a second Bayer noise process to obtain a third processed image, wherein precision of the second Bayer noise process is lower than precision of the first Bayer noise process; Mapping the privacy area marked in the second processed image to the third processed image to obtain a fourth processed image; The fourth processed image is input into a privacy grading model to obtain a comprehensive privacy level score for each privacy area.

7. The method according to claim 1, wherein Based on the comprehensive privacy level score of each privacy area, the corresponding desensitization method is determined to perform pixel-level privacy processing, including: Determining a corresponding privacy level based on the comprehensive privacy level scores of the privacy areas; Determine corresponding desensitization measures based on the privacy level; Based on corresponding desensitization measures, pixel-level privacy processing is performed on the second processed image.

8. The method according to claim 1, wherein The desensitization means include: Mild blurring, pixelation, non-decipherable mosaic overlays, and decipherable mosaic overlays of private areas.

9. A device for desensitizing image data, characterized in that: include: A first acquisition module is used to acquire a RAW domain image of an image to be processed or a video stream to be processed; an image processing module, configured to sequentially perform wide dynamic range processing, first Bayer noise processing, and tone mapping processing on the RAW domain image to obtain a first processed image; The second acquisition module is used to obtain a dynamic ROI detection model and a privacy classification model; a marking module, configured to output a second processed image based on the first processed image and the dynamic ROI detection model, wherein the second processed image is a privacy area marked on the first processed image; an obtaining module, configured to obtain a comprehensive privacy level score for each privacy area based on the second processed image and the privacy grading model; The desensitization processing module is used to determine the corresponding desensitization means based on the comprehensive privacy level score of each privacy area to perform pixel-level privacy processing.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

Images