Image desensitization processing method and device

Abstract

The embodiment of the invention provides an image desensitization processing method and device. The method comprises the steps that a to-be-desensitized image and structured data indicated for the image are determined; inputting the image and the structured data indicated for the image into a preset recognition model framework, and recognizing an object of a target type in the image through a target recognition model in the preset recognition model framework to obtain a sensitive object in the image; and desensitizing the image according to the sensitive object. The method is used for achieving the effect of accurately and effectively identifying the sensitive information in the image.

Description

Technical Field

[0001] The present application relates to the field of image processing technology and artificial intelligence, and in particular to a method and device for desensitizing an image. Background Art

[0002] After capturing an image, it is often necessary to detect sensitive objects in the image (such as faces, license plates, etc.) and then desensitize the image.

[0003] However, when desensitizing images, it is necessary to set up corresponding image desensitization processing mechanisms separately for different sensitive objects. For example, when desensitizing images containing faces and license plates, professionals are required to design corresponding image desensitization processing mechanisms based on the different features of faces and license plates respectively. In addition, a large number of images generally contain a variety of sensitive objects. On the one hand, technical personnel are required to have a high level of professional expertise to design corresponding image desensitization processing mechanisms for different types of sensitive objects. On the other hand, when desensitizing a large number of images, users need to find corresponding image desensitization processing mechanisms for different images for processing.

[0004] However, the above manual processing method requires a lot of manpower and time, which leads to low efficiency of image desensitization and requires a lot of manpower costs. Summary of the invention

[0005] The embodiments of the present application provide a method and device for desensitizing an image, so as to achieve the effect of quickly identifying the desensitization of an image.

[0006] In a first aspect, an embodiment of the present application provides a method for desensitizing an image, comprising:

[0007] Determining an image to be desensitized and structured data indicated for the image;

[0008] The image and the structured data indicated by the image are input into a preset recognition model framework, and the object of the target type in the image is recognized by the target recognition model in the preset recognition model framework to obtain the sensitive object in the image; wherein the preset recognition model framework includes sensitive object recognition models corresponding to different object types, the target recognition model is a sensitive object recognition model matched with the structured data indicated by the image, and the target type is the object type corresponding to the target recognition model;

[0009] A desensitization operation is performed on the image according to the sensitive object.

[0010] In a possible implementation, determining an image to be desensitized and structured data indicated for the image includes:

[0011] Parsing the first communication protocol message to obtain the image and structured data indicated for the image; or

[0012] An image is received, and a second communication protocol message for the image is parsed to obtain structured data for the image.

[0013] In a possible implementation manner, the first communication protocol message is sent by the client, or the first communication protocol message is determined based on a first information indication operation triggered by a human-computer interaction interface; and / or

[0014] The second communication protocol message is sent by the client, or the second communication protocol message is determined based on a second information indication operation triggered through a human-computer interaction interface.

[0015] In one possible implementation, the structured data includes a service address of the target recognition model, an object type corresponding to the target recognition model, a model version of the target recognition model, a confidence threshold of the target recognition model, a review mark of the image area recognized by the target recognition model, a batch size of the target recognition model, and at least one information in the preprocessed image.

[0016] In a possible implementation, performing a desensitization operation on an image according to a sensitive object includes:

[0017] The sensitive object is sent to the desensitizing end so that the desensitizing end performs a desensitizing operation on the image according to the sensitive object.

[0018] In a possible implementation, the structured data includes a method for indicating an object type corresponding to the target recognition model; before identifying an object of the target type in the image by the target recognition model in the preset recognition model framework to obtain a sensitive object in the image, the method further includes:

[0019] The sensitive object recognition model corresponding to the object type indicated by the structured data in the preset recognition model framework is determined as the target recognition model.

[0020] In a possible implementation, identifying the target type object in the image by using the target recognition model in the preset recognition model framework to obtain the sensitive object in the image includes:

[0021] The structured data includes a confidence threshold for indicating a target recognition model, the image is input into the target recognition model, and each candidate sensitive object belonging to the target type in the image and the confidence of each candidate sensitive object are obtained; the candidate sensitive object whose confidence is greater than or equal to the confidence threshold is determined as the sensitive object in the image output by the target recognition model; and / or

[0022] The structured data includes a re-inspection mark of the image area identified by the target recognition model, and the image is input into the target recognition model; in response to the re-inspection mark indicating that re-inspection is required, sensitive object areas in the image are detected based on the target recognition model, and second-order detection processing is performed on each of the sensitive object areas to obtain sensitive objects in the image output by the target recognition model; in response to the re-inspection mark indicating that re-inspection is not required, the sensitive object areas in the image obtained are determined based on the target recognition model as the sensitive objects in the image output by the target recognition model; wherein the sensitive object areas are areas in the image where sensitive objects belonging to the target type are located.

[0023] In a possible implementation, before the image and the structured data indicated for the image are input into a preset recognition model framework, the method further includes: based on a preset load balancing strategy, selecting the preset recognition model framework from at least two initial recognition model frameworks, each of the initial recognition model frameworks including at least one sensitive object recognition model; and / or

[0024] Before the target type object in the image is identified by the target recognition model in the preset recognition model framework to obtain the sensitive object in the image, it also includes: performing a preprocessing operation corresponding to the target recognition model on the image according to the preprocessing requirements of the target recognition model; the preprocessing operation, the desensitization processing operation, and the request to call the target recognition model are encapsulated in the same software toolkit.

[0025] In a second aspect, an embodiment of the present application provides a method for desensitizing an image, comprising:

[0026] Receive an image and a sensitive object in the image sent by a server; the sensitive object is obtained by the server executing the first aspect and / or various possible implementation methods of the first aspect as described above.

[0027] Desensitize images based on sensitive objects.

[0028] In a third aspect, an embodiment of the present application provides a method for desensitizing an image, comprising:

[0029] An image to be desensitized and structured data indicated for the image are determined.

[0030] The image and structured data indicated by the image are sent to the server, so that the server obtains the sensitive objects in the image through the first aspect and / or various possible implementation methods of the first aspect as described above.

[0031] Receive sensitive objects in the image returned by the server, and perform a desensitization operation on the image according to the sensitive objects; or, receive a desensitized image corresponding to the image, where the desensitized image is obtained by performing a desensitization operation on the image according to the sensitive objects.

[0032] In a possible implementation, sending the image and structured data indicated by the image to the server includes:

[0033] encapsulating the image and the structured data indicated by the image into a first communication protocol message, and sending the first communication protocol message to the server; or

[0034] The structured data indicated for the image is encapsulated into a second communication protocol message, and the image and the second communication protocol message for the image are sent to the server.

[0035] In a possible implementation, the first communication protocol message is determined based on a first information indication operation triggered through a human-computer interaction interface; and / or

[0036] The second communication protocol message is determined based on a second information indication operation triggered through the human-computer interaction interface.

[0037] In one possible implementation, the structured data includes a service address of the target recognition model, an object type corresponding to the target recognition model, a model version of the target recognition model, a confidence threshold of the target recognition model, a review mark of the image area recognized by the target recognition model, a batch size of the target recognition model, and at least one information in the preprocessed image.

[0038] In a possible implementation, performing a desensitization operation on an image according to a sensitive object includes:

[0039] At least one of marking, blurring, encryption and occlusion processing is performed on sensitive objects in the image to obtain a desensitized image.

[0040] In a fourth aspect, an embodiment of the present application provides a computer device, including: a memory and a processor.

[0041] The memory stores computer-executable instructions.

[0042] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0043] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementations of the first aspect.

[0044] The embodiments of the present application provide a method and device for desensitizing an image. Using each desensitized object recognition model in a preset recognition model framework, images can be recognized in parallel, thereby obtaining sensitive objects under multiple object types, and then each sensitive object in the image can be quickly identified, further facilitating continued desensitization based on the obtained sensitive objects. Especially when processing a large number of images, based on each desensitized object recognition model in the preset recognition model framework provided by this embodiment, the recognition of sensitive objects in a large number of images can be quickly completed, thereby improving the image recognition efficiency; further, when a large number of images need to be desensitized, the desensitization process of a large number of images can be quickly completed, which can improve the image desensitization efficiency.

[0045] Furthermore, there is no need to activate all sensitive object recognition models in the preset recognition model framework to perform recognition and processing on the image, thereby improving the processing efficiency of the system and providing users with greater autonomy in operating computer devices, so that they can freely select and process the types of sensitive objects of interest; and there is no need to rely on the user's professional processing knowledge to select the sensitive object recognition model to be used. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0047] Figure 1 The present invention provides a schematic diagram of a structure of an inference service module deployed in a computer device. Figure 1 ;

[0048] Figure 2 A schematic diagram of a structure of an inference service module deployed in a computer device provided in an embodiment of the present application Figure 2 ;

[0049] Figure 3 A schematic diagram of a structure of an inference service module deployed in a computer device provided in an embodiment of the present application Figure 3 ;

[0050] Figure 4A flowchart of a method for training a sensitive object recognition model provided in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 1 ;

[0052] Figure 6 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 2 ;

[0053] Figure 7 Interface diagram provided for the embodiment of the present application;

[0054] Figure 8 A schematic diagram of the structure of a computer device provided in an embodiment of the present application Figure 1 ;

[0055] Fig. 9 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 3 ;

[0056] Fig.10 A schematic diagram of the structure of a computer device provided in an embodiment of the present application Figure 2 ;

[0057] Fig.11 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 4 ;

[0058] Fig.12 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 1 ;

[0059] Fig.13 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 2 ;

[0060] Fig.14 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 3 ;

[0061] Fig.15 A schematic diagram of the structure of a computer device provided in an embodiment of the present application.

[0062] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0063] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0064] In the digital age, it is often necessary to collect various images, identify and analyze the content in the images, and obtain various processing results of the target. In addition to the content related to the processing results, these images also contain various sensitive objects. For example, sensitive objects include privacy data, facial images, and license plate numbers. These sensitive objects contain privacy data.

[0065] For example, in the field of autonomous driving of vehicles, in order to provide support for autonomous driving of vehicles, it is necessary to collect map data, and various sensitive objects are widely present in these map data, such as faces, license plates, and so on.

[0066] In order to avoid security issues caused by privacy leakage, it is necessary to desensitize the sensitive objects in the collected images, and then identify and analyze the desensitized images; furthermore, it is necessary to identify the sensitive objects in the images, and then desensitize the sensitive objects in the images based on the identified sensitive objects.

[0067] In one example, a technician can identify different types of sensitive objects in an image and then perform desensitization on the image. However, manually completing the image desensitization operation requires the professional knowledge of the technician, takes a lot of time, and results in low processing efficiency. It is also prone to errors, resulting in errors in the image desensitization operation.

[0068] The present application provides an image desensitization processing method and device, aiming to solve the above technical problems of the prior art.

[0069] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0070] Figure 1 The present invention provides a schematic diagram of a structure of an inference service module deployed in a computer device. Figure 1 ,like Figure 1As shown, the reasoning service module includes multiple sensitive object recognition models; different sensitive object recognition models correspond to different object types, and each sensitive object recognition model is based on the same deep learning framework.

[0071] The sensitive object recognition model is used to recognize images and obtain sensitive objects under the object type corresponding to the sensitive object recognition model.

[0072] In one example, each sensitive object recognition model is encapsulated in the same reasoning service module.

[0073] Exemplarily, a reasoning service module may be deployed in a computer device, wherein the reasoning service module includes a plurality of sensitive object recognition models.

[0074] The computer device may be, for example, any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., without limitation.

[0075] like Figure 1 As shown, the reasoning service module includes the first sensitive object recognition model, the second sensitive object recognition model, the third sensitive object recognition model, ..., the i-th sensitive object recognition model, ..., the N-th sensitive object recognition model. N is a positive integer greater than 1; i is a positive integer greater than or equal to, less than or equal to N.

[0076] Different sensitive object recognition models correspond to different object types. The sensitive object recognition model is used to recognize an image, and then obtain a sensitive object under the object type corresponding to the sensitive object recognition model.

[0077] The object type includes any one or more combinations of the following: face object type, license plate object type, preset identification object type, preset area object type, moving target object type, and the like.

[0078] For example, the preset area object is a sensitive geographical area type, and the moving target object is a pedestrian object type or other object types in motion.

[0079] In one example, the preset recognition model framework can be obtained based on the same deep learning framework. For example, the deep learning framework can be any one of the PyTorch framework, the TensorFlow framework, the Caffe framework, and the like.

[0080] Among them, each sensitive object recognition model is encapsulated in the same reasoning service module, which is more conducive to parallel image processing, convenient for staff to uniformly perform version control, update and maintenance, and reduces the complexity of management.

[0081] In this embodiment, by providing a reasoning service module with multiple sensitive object recognition models deployed, wherein different sensitive object recognition models correspond to different object types; each sensitive object recognition model is used to recognize an image, and obtain a sensitive object under the object type corresponding to the sensitive object recognition model; and then using the reasoning service module, the image can be recognized in parallel, and then sensitive objects under multiple object types can be obtained, and then each sensitive object in the image can be quickly identified, and further facilitates the continued desensitization processing based on the obtained sensitive objects. Especially when processing a large number of images, the reasoning service module with multiple sensitive object recognition models deployed based on this embodiment can quickly complete the recognition of sensitive objects in a large number of images, thereby improving the image recognition efficiency; further, when a large number of images need to be desensitized, the desensitization processing process for a large number of images can be quickly completed, which can improve the image desensitization efficiency. Moreover, when identifying sensitive objects in an image, the required sensitive object recognition model can be selected from the reasoning service module, without relying on the user's professional processing knowledge.

[0082] Moreover, since each sensitive object recognition model can rely on the same operating environment under the same deep learning framework, when it is necessary to add or delete sensitive objects that need to be recognized by the inference service module, that is, when the number of sensitive object recognition models is adjusted, the newly added or reduced sensitive object recognition model will not affect the existing sensitive object recognition models in the inference service module, thereby reducing the tediousness of developers' adjustments to the inference service module and alleviating the workload of developers.

[0083] In one example, different sensitive object recognition models correspond to different preprocessing operations.

[0084] Exemplarily, the sensitive object recognition model corresponds to a preprocessing operation. For example, there is a one-to-one correspondence between the sensitive object recognition model and the preprocessing operation. For another example, the same sensitive object recognition model corresponds to multiple preprocessing operations, but the preprocessing operations of different sensitive object recognition models are different. For another example, the same sensitive object recognition model corresponds to multiple preprocessing operations, but different sensitive object recognition models may have the same preprocessing operations.

[0085] The preprocessing operation includes any of the following:

[0086] If it is determined that the resolution of the image is greater than a first preset value, the image is reduced.

[0087] If it is determined that the resolution of the image is less than a second preset value, the image is enlarged and interpolated; wherein the second preset value is less than the first preset value.

[0088] If it is determined that the size of the image is smaller than the preset size, a to-be-filled area in the image is determined, and a filling process is performed on the to-be-filled area.

[0089] Among them, the first preset value, the second preset value, and the preset size may be empirical values.

[0090] Different preprocessing operations are used for different sensitive object recognition models. This is because different types of sensitive objects have different attributes. For example, pedestrian objects, vehicle objects, and sensitive geographic area objects are larger in size and have richer features than face objects, license plate objects, and traffic sign objects. Therefore, different image preprocessing operations are required for different types of sensitive objects.

[0091] Among them, the preprocessing operation can be deployed to other computer devices, for example, the preprocessing operation can be deployed to the preprocessing module, and the preprocessing module can be deployed to the preprocessing end, which can be any of the following: terminal equipment, independent server, server cluster, cloud, Internet of Things, Internet of Vehicles, distributed system, etc., without limitation.

[0092] Alternatively, the preprocessing operations can be deployed together with the inference serving module on the same computer device. Figure 2 A schematic diagram of a structure of an inference service module deployed in a computer device provided in an embodiment of the present application Figure 2 ,like Figure 2 As shown, the preprocessing operation is deployed to the preprocessing module, and the preprocessing module and the reasoning service module are deployed together in the same computer device.

[0093] In one example, a desensitization operation needs to be deployed.

[0094] For example, after the sensitive object recognition model is used to recognize the image, the sensitive object under the object type corresponding to the sensitive object recognition model can be obtained. After the sensitive object is obtained, the sensitive object in the image can also be desensitized.

[0095] Among them, the desensitization processing operation includes any one or more combinations of the following: marking processing, blurring processing, encryption processing, and blocking processing.

[0096] Among them, the desensitizing processing operation can be deployed to other computer devices. For example, the desensitizing processing operation can be deployed to the desensitizing processing module, and the desensitizing processing module can be deployed to the desensitizing end. The desensitizing end can be any of the following: terminal equipment, independent server, server cluster, cloud, Internet of Things, Internet of Vehicles, distributed system, etc., without any restriction.

[0097] Alternatively, the desensitization operation can be deployed together with the reasoning service module on the same computer device. Figure 3 A schematic diagram of a structure of an inference service module deployed in a computer device provided in an embodiment of the present application Figure 3 ,like Figure 3 As shown, the desensitizing processing operation is deployed to the desensitizing processing module, and the desensitizing processing module and the reasoning service module are deployed together in the same computer device.

[0098] In one example, the preprocessing operation, the desensitization operation on the image, and the request to call the sensitive object recognition model are encapsulated in the same software toolkit.

[0099] Exemplarily, in order to facilitate the calling of the above-mentioned preprocessing operation and the above-mentioned desensitizing processing operation, and to facilitate the calling of the sensitive object recognition model, the above-mentioned preprocessing operation, the above-mentioned desensitizing processing operation, and the request for calling the sensitive object recognition model can be encapsulated in the same software toolkit (Software Development Kit, SDK).

[0100] Figure 4 A flowchart of a method for training a sensitive object recognition model provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, the method includes:

[0101] S401, training an initial model based on images.

[0102] Exemplarily, the execution subject of the embodiment of the present application may be a computer device, which is not limited to this. The computer device may be any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., which is not limited to this.

[0103] The computer device obtains images to be trained; wherein, these images to be trained can be obtained by continuously shooting the surrounding environment of the vehicle based on the shooting equipment deployed in the vehicle.

[0104] Among them, a reasoning service module is deployed in the computer device, and the reasoning service module includes multiple initial models, and the initial models correspond to the object types.

[0105] The computer device inputs the acquired images to be trained into each initial model in the reasoning service module for processing, so as to train each initial model.

[0106] Among them, the computer device can input the image to be trained into each initial model, and then output the sensitive object recognition result under the object type corresponding to the initial model based on the initial model; then, the computer device adjusts the parameters in the initial model according to the sensitive object recognition result until the accuracy of the sensitive object recognition result output by each initial model reaches a predetermined accuracy.

[0107] Alternatively, when only part of the initial model in the inference service module needs to be trained, the computer device provides an indication information indicating the initial model that needs to be trained. Then, the computer device can input the image to be trained into the initial model that needs to be trained, and then output the sensitive object recognition result under the object type corresponding to the initial model based on the initial model; then, the computer device adjusts the parameters in the initial model according to the sensitive object recognition result until the accuracy of the sensitive object recognition result output by the initial model that needs to be trained reaches a predetermined accuracy.

[0108] S402. Obtain a sensitive object recognition model corresponding to the initial model; wherein different initial models correspond to different object types, and each initial model is based on the same deep learning framework.

[0109] The sensitive object recognition model is used to recognize images and obtain sensitive object recognition results under the object type corresponding to the sensitive object recognition model.

[0110] In one example, each sensitive object recognition model is encapsulated in the same reasoning service module.

[0111] In one example, different sensitive object recognition models correspond to different preprocessing operations; the preprocessing operation, the desensitization operation, and the request to call the target recognition model are encapsulated in the same software toolkit.

[0112] Exemplarily, through the previous steps, after the model training is completed, a sensitive object recognition model corresponding to the initial model is obtained. Each sensitive object recognition model is trained based on the same deep learning framework. The computer device encapsulates the obtained sensitive object recognition models in the same reasoning service module.

[0113] Among them, each sensitive object recognition model is used to recognize the image, and then obtain the sensitive object recognition result under each object type.

[0114] The embodiment of the present application provides a method for training a sensitive object recognition model, which can quickly complete the training of multiple sensitive object recognition models by inputting the images to be trained into different sensitive object recognition models for parallel training. Among them, each sensitive object recognition model is based on the same deep learning framework, and since each sensitive object recognition model is trained based on a deep learning framework, the required operating environment is the same; therefore, under the condition that a new object type needs to be added, a sensitive object recognition model suitable for this framework can be trained again based on this deep learning framework without reconfiguring the operating environment. In this way, the coupling between the various sensitive object recognition models in the system is reduced, and the assembly flexibility between the various sensitive object recognition models is improved.

[0115] In addition, since the underlying training logic of the sensitive object recognition model is the same, it only needs to be adaptively adjusted during the training process for different object types, thereby reducing the training burden of the sensitive object recognition model; in addition, each sensitive object recognition model is trained based on the same deep learning framework, and there is no need to make large-scale adjustments to the allocation of device resources, thereby reducing the workload of developers.

[0116] Figure 5 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 1 ,like Figure 5 As shown, the method includes:

[0117] S501: Determine an image to be desensitized and structured data indicated by the image.

[0118] For example, the embodiments of the present application are described by taking the execution subject as a computer device as an example. The computer device is, for example, any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., without limitation.

[0119] The following is an introduction to the map collection scenario of the vehicle's autonomous driving, which should not be understood as a limitation on the application scenarios of the embodiments of the present application. For example, it can also be the desensitization processing of collected medical imaging data (e.g., computed tomography, magnetic resonance imaging), images uploaded by users on social media platforms, and surveillance images collected by security monitoring systems.

[0120] First, the surrounding environment is photographed by an image capturing device installed in the vehicle to form an image, which is then stored in a computer device.

[0121] For example, the image capturing device carried in the vehicle may be a common camera, a fisheye camera, or a surround view camera. Among them, the content captured by some image capturing devices includes video stream data. In this case, the computer device extracts frames from the video stream data to obtain one or more images, which are then combined with the one or more images obtained by the capture to obtain each image.

[0122] The frame extraction process can extract image frames from the video stream data at fixed frame intervals (for example, every 3 frames, every 5 frames), thereby obtaining one or more images.

[0123] The image may include: road traffic conditions, road vehicle conditions, road pedestrian conditions, and surrounding buildings and area conditions.

[0124] For example, road traffic conditions include: traffic signals, traffic signs (e.g., height limit signs, speed limit signs, give way signs); road vehicle conditions include: whether there are certain types of vehicles, for example, the types of vehicles may include cars, trucks, motorcycles, and buses; road pedestrian conditions include: the location of pedestrians around the vehicle; surrounding buildings and areas include: buildings, gates, and fences around the vehicle. Next, the computer device acquires and determines the image to be desensitized and the structured data corresponding to the image; wherein the structured data is indicated by the image and can be input by the user and transmitted to the computer device.

[0125] S502. Input the image and the structured data indicated by the image into a preset recognition model framework, and identify the target type object in the image through the target recognition model in the preset recognition model framework to obtain the sensitive object in the image; wherein, the preset recognition model framework includes sensitive object recognition models corresponding to different object types, the target recognition model is a sensitive object recognition model matched with the structured data indicated by the image, and the target type is the object type corresponding to the target recognition model.

[0126] Exemplarily, a preset recognition model framework is deployed in a computer device, wherein the preset recognition model framework includes a plurality of sensitive object recognition models; different sensitive object recognition models correspond to different object types.

[0127] Among them, the preset recognition model framework is as shown above Figure 1-Figure 3 The reasoning service module in the preset recognition model framework is shown in Figure 1-Figure 3 The introduction of , no longer repeat.

[0128] The preset recognition model framework is based on Figure 4 The training process is obtained by the training method shown in Figure 4 The introduction of , no longer repeat.

[0129] The computer device inputs the image to be desensitized and the structured data into the above-mentioned preset recognition model framework. The computer device retrieves the sensitive object recognition model matched by the structured data according to the structured data; the computer device determines that the retrieved sensitive object recognition model is the current target recognition model. Then, the computer device recognizes the image to be desensitized based on the current target recognition model to obtain the sensitive object in the image; at this time, the object type of the obtained sensitive object corresponds to the current target recognition model.

[0130] When the structured data indicates multiple sensitive object recognition models, the computer device simultaneously determines multiple target recognition models; then, the computer device executes the multiple target recognition models in parallel, and recognizes images based on the multiple target recognition models in parallel to obtain the sensitive objects of the image output by each target recognition model.

[0131] The object type includes any one or more combinations of the following: face object type, license plate object type, preset identification object type, preset area object type, moving target object type, and the like.

[0132] Furthermore, sensitive objects include any one or more combinations of the following: face objects, license plate objects, preset identification objects, preset area objects, moving target objects, and the like.

[0133] S503: Perform a desensitization operation on the image according to the sensitive object.

[0134] Exemplarily, the computer device performs a desensitization operation on the image according to each sensitive object identified by each sensitive object recognition model.

[0135] Among them, the desensitization processing operation includes any one or more combinations of the following: marking processing, blurring processing, encryption processing, and blocking processing.

[0136] For example, the computer device performs mosaic processing on the sensitive objects identified in the image. Thus, the desensitized image can be put into subsequent use without leaking privacy, such as data analysis for autonomous driving of vehicles.

[0137] In this embodiment, by enabling the sensitive object recognition model corresponding to the type of sensitive object that the user is interested in according to the type of sensitive object indicated in the structured data, the image is processed to obtain each sensitive object in the image, and then desensitization processing can be performed on each identified sensitive object. In this way, there is no need to enable all sensitive object recognition models in the preset recognition model framework to recognize and process the image, thereby improving the processing efficiency of the system, and providing users with greater autonomy in controlling computer devices, so that they can freely select and process the type of sensitive objects of interest; and there is no need to rely on the user's professional processing knowledge to select the sensitive object recognition model to be used.

[0138] Moreover, by using each desensitized object recognition model in the preset recognition model framework, images can be recognized in parallel, and sensitive objects under multiple object types can be obtained, and each sensitive object in the image can be quickly identified, which further facilitates the continued desensitization processing based on the obtained sensitive objects. Especially when processing a large number of images, based on each desensitized object recognition model in the preset recognition model framework provided in this embodiment, the recognition of sensitive objects in a large number of images can be quickly completed, thereby improving the image recognition efficiency; further, when a large number of images need to be desensitized, the desensitization processing process for a large number of images can be quickly completed, which can improve the image desensitization efficiency.

[0139] In addition, the various sensitive object recognition models provided in the embodiments of the present application are based on the same deep learning framework. Therefore, developers do not need to use different deep learning frameworks to train each sensitive object recognition model for each sensitive object type. Furthermore, the reasoning service module only includes one deep learning framework, and there is no need to set up multiple deep learning frameworks, nor is there any need to adjust and set up the operating environment for multiple deep learning frameworks, thereby reducing the workload of developers.

[0140] Figure 6 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 2 ,like Figure 6 As shown, the method includes:

[0141] S601. According to the preprocessing requirements of the target recognition model, a preprocessing operation corresponding to the target recognition model is performed on the image.

[0142] In one example, a preprocessing operation, a desensitizing operation, and a request to retrieve a target recognition model are encapsulated in the same software toolkit.

[0143] For example, the embodiments of the present application are described by taking the execution subject as a computer device as an example. The computer device is, for example, any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., without limitation.

[0144] In one example, the preprocessing operation includes any of the following:

[0145] If it is determined that the resolution of the image is greater than a first preset value, the image is reduced.

[0146] If it is determined that the resolution of the image is less than a second preset value, the image is enlarged and interpolated; wherein the second preset value is less than the first preset value.

[0147] If it is determined that the size of the image is smaller than the preset size, a to-be-filled area in the image is determined, and a filling process is performed on the to-be-filled area.

[0148] Due to the differences in the properties of sensitive objects of different object types and the processing properties of target recognition models corresponding to different object types, different preprocessing operations are performed on the images, and the different preprocessing operations are adapted to the preprocessing requirements of different target recognition models.

[0149] In this way, the preprocessed image is more in line with the processing requirements of each target recognition model, thereby improving the accuracy of the target recognition model in identifying sensitive objects.

[0150] For example, since a vehicle is often equipped with multiple image capture devices, and there are differences between each image capture device, the image quality of images captured by each image capture device is also different. For example, when an image includes multiple pictures, the resolutions of pictures captured by different image capture devices are different.

[0151] In some cases, for images with a resolution greater than a first preset value, the computer device performs downscaling processing to reduce the occupancy of the memory (Random Access Memory, RAM) or video memory (Video Random Access Memory, also referred to as video memory) in the computer device, thereby improving the processing efficiency of the computer device.

[0152] For example, the first preset value may be 3840 pixels×2160 pixels (corresponding to a 4K image), or the first preset value may be 2048 pixels×1080 pixels (corresponding to a 2K image). It is worth noting that this embodiment does not limit the value of the first preset value.

[0153] In other cases, for images with a resolution less than the second preset value, the computer device enlarges and interpolates them to ensure that the content in the image can be recognized by the desensitized object recognition model, thereby ensuring the recognition accuracy of the desensitized object recognition model.

[0154] For example, the second preset value may be 100 pixels×100 pixels. It is worth noting that this embodiment does not limit the value of the second preset value.

[0155] After performing corresponding preprocessing on images with a resolution greater than the first preset value and images with a resolution less than the second preset value, it can be ensured that the resolution of each picture in the image is between the second preset value and the first preset value; therefore, the computational burden of the computer equipment can be reduced while ensuring that the desensitized object recognition model can recognize various objects in the image, thereby improving processing efficiency.

[0156] In addition, in order to facilitate the desensitized object recognition model to identify sensitive objects in the image, the size of each image can also be normalized.

[0157] If the computer device determines that the size of the image is smaller than the preset size, the area to be filled is determined based on the difference between the size of the image and the preset size; then, the image is filled based on the area to be filled so that the image meets the preset size, so as to facilitate the desensitized object recognition model to identify sensitive objects in the image.

[0158] For example, the preset size indicates that the length dimension and width dimension of the image are equal (that is, the image is required to be a square), while the length dimension and width dimension of the preprocessed image are not equal (that is, the input image is a rectangle); the computer device determines the area to be filled based on the difference between the length dimension and width dimension of the preprocessed image, so that the long side of the image is equal to the short side.

[0159] For example, the length dimension of the preprocessed image is 300 pixels, and the width dimension is 150 pixels. The difference between the length dimension and the width dimension of the preprocessed image is 150 pixels, and it is determined that the width direction of the preprocessed image needs to be filled; next, the computer device uses a pure gray (RGB value is 114, 114, 114) color block to fill a filling area with a width of 150 pixels in the width direction of the preprocessed image; thus, the length dimension of the preprocessed image is 300 pixels, and the width dimension is 300 pixels, and the length dimension is equal to the width dimension, which is convenient for subsequent processing of the desensitized object recognition model, thereby improving the recognition accuracy of the desensitized object recognition model.

[0160] On the other hand, desensitization processing includes one or more of the following: marking processing, blurring processing, encryption processing, and occlusion processing; wherein, marking processing represents drawing the outer frame of the sensitive object; blurring processing represents blurring the sensitive object; encryption processing represents processing the sensitive object based on an encryption algorithm; occlusion processing represents occluding the sensitive object based on preset color blocks.

[0161] For example, image desensitization processing can be performed through marking processing, blurring processing, encryption processing, or occlusion processing.

[0162] The labeling process refers to drawing the outer frame of the sensitive object, that is, drawing the outer frame of each sensitive object through the predicted frame of the sensitive object.

[0163] Blurring refers to blurring sensitive objects through mosaics, random noise, etc., making them unrecognizable to the human eye, thereby achieving the effect of desensitization.

[0164] Encryption processing refers to encrypting the area where each sensitive object is located based on an encryption algorithm.

[0165] Occlusion processing refers to using preset color blocks to occlude the area where each sensitive object is located. For example, non-sensitive objects in the vicinity of each sensitive object can be copied and the sensitive objects can be occluded using the copied non-sensitive objects. Alternatively, color blocks with certain preset colors can be used to occlude the area where the sensitive objects are located.

[0166] S602: Determine an image to be desensitized and structured data indicated for the image.

[0167] Exemplarily, the specific process of this step can refer to the content in S301, which will not be repeated here.

[0168] In one example, step S602 may include the following process:

[0169] Parse the first communication protocol message to obtain the image and structured data indicated for the image; or receive the image and parse the second communication protocol message for the image to obtain structured data for the image.

[0170] Exemplarily, in one case, the structured data and the preprocessed image are encapsulated to obtain a first communication protocol message, that is, the first communication protocol message includes the image and the structured data indicated for the image.

[0171] In another case, the second communication protocol message is encapsulated and the computer device receives the pre-processed image.

[0172] The first communication protocol message and the second communication protocol message may be applicable to any network communication protocol, for example, may be an http request (HyperText Transfer Protocol Request, hypertext transfer protocol request, referred to as http request).

[0173] The computer device sends the first communication protocol message or the second communication protocol message to the inference service module, and based on the first communication protocol message or the second communication protocol message, calls the corresponding desensitized object recognition model to further identify the sensitive objects in the image.

[0174] In one example, the first communication protocol message is sent by the client, or the first communication protocol message is determined based on a first information indication operation triggered by a human-computer interaction interface; and / or the second communication protocol message is sent by the client, or the second communication protocol message is determined based on a second information indication operation triggered by the human-computer interaction interface.

[0175] Exemplarily, in one case, the first communication protocol message may be sent by the client. For example, after the client uploads the image and inputs the structured data, the client encapsulates the image and the structured data into the first communication protocol message and sends it to the computer device.

[0176] In another case, the first communication protocol message may be determined by a first information indication operation triggered by a user based on a human-computer interaction interface. Figure 7 The interface diagram provided for the embodiment of the present application is as follows: Figure 7 As shown in (a), a human-computer interaction interface is provided in a client (such as a web page, a mobile application or other software interface), and the human-computer interaction interface includes a selection button. Figure 7 As shown in (b), after the user triggers the selection button on the human-computer interaction interface, the client displays a pop-up window. Then, the user selects or uploads an image through the pop-up window, and the user enters structured data through the first indication information triggered by the human-computer interaction interface. For example, the user can upload images and enter structured data by clicking other buttons, selecting files, filling out forms, etc. Among them, the first indication information may include: 1) Instructing the user to upload an image: the user selects or drags an image file to the upload area. 2) Instructing to submit a form: After the user fills out a form containing structured data, click the "Submit" button. 3) Instructing the user to select a file: the user selects the image file to be uploaded by clicking the "Browse" or "Select File" button.

[0177] After the client uploads the image and inputs the structured data, the client encapsulates the image and the structured data into a first communication protocol message and sends it to the computer device.

[0178] In another case, the second communication protocol message is sent by the client. For example, after the client uploads the image and inputs the structured data, the client encapsulates the structured data into the second communication protocol message and sends the image and the second communication protocol message to the computer device.

[0179] In another case, the second communication protocol message is determined based on a second information indication operation triggered through the human-computer interaction interface.

[0180] For example, Figure 7 As shown in (a), a human-computer interaction interface is provided in a client (such as a web page, a mobile application or other software interface), and the human-computer interaction interface includes a selection button. Figure 7 As shown in (b), after the user triggers the selection button on the human-computer interaction interface, the client displays a pop-up window. Then, the user selects or uploads an image through the pop-up window, and the user enters structured data through the second indication information triggered by the human-computer interaction interface. For example, the user can fill in and upload structured data through the second indication information by clicking a button, triggering a screen, etc. For example, the second indication information may include: 1) Instructing the user to upload an image: the user selects or drags an image file to the upload area. 2) Instructing to submit a form: After the user fills in a form containing structured data, clicks the "Submit" button.

[0181] After the client uploads the image and inputs the structured data, the client encapsulates the structured data into a second communication protocol message and sends the image and the second communication protocol message to the computer device.

[0182] In one example, the structured data includes a service address of the target recognition model, an object type corresponding to the target recognition model, a model version of the target recognition model, a confidence threshold of the target recognition model, a review mark of the image area recognized by the target recognition model, a batch size of the target recognition model, and at least one information in the preprocessed image.

[0183] In one example, the structured data includes a confidence threshold for indicating a target recognition model.

[0184] The target type objects in the image are identified by the target recognition model in the preset recognition model framework to obtain sensitive objects in the image, including:

[0185] The image is input into the target recognition model to obtain candidate sensitive objects belonging to the target type in the image and the confidence of each candidate sensitive object.

[0186] Candidate sensitive objects whose confidence is greater than or equal to the confidence threshold are determined as sensitive objects in the image output by the target recognition model.

[0187] Exemplarily, the structured data includes the service address of the desensitized object recognition model, the object type of each desensitized object recognition model, the model version of each desensitized object recognition model, the confidence threshold of each desensitized object recognition model, the review flag of each desensitized object recognition model, and the batch size of each desensitized object recognition model.

[0188] For example, a plurality of inference service modules (also referred to as inference service nodes or triton service nodes) and a load balancing module may be deployed in a computer device; wherein a load balancing strategy is pre-stored in the load balancing module.

[0189] The computer device sends the first communication protocol message or the second communication protocol message to the load balancing module. Since the first communication protocol message or the second communication protocol message contains structured data, the structured data includes the service address of the sensitive object identification model sent to the load balancing module. The load balancing module sends the service address of the sensitive object identification model to the corresponding reasoning service module according to the load balancing strategy; wherein each reasoning service module corresponds to a node address, and the load balancing module can call the reasoning service module corresponding to the node address through each node address, thereby starting the reasoning service module for processing.

[0190] By installing multiple reasoning service modules in a computer device, the corresponding reasoning service module can be selected and started according to the load balancing strategy to process the image to meet the high concurrency demand of image desensitization. More images can be processed in the same time, thereby improving processing efficiency. The object type of each desensitized object recognition model includes detection parameters corresponding to each object type. The value of each detection parameter represents whether the desensitized object recognition model detects the object type corresponding to this detection parameter.

[0191] For example, if the value of the detection parameter is 1, it indicates that the object type corresponding to the detection parameter needs to be detected; if the value of the detection parameter is 0, it indicates that the object type corresponding to the detection parameter does not need to be detected.

[0192] Each activated desensitized object recognition model processes the image, identifies the sensitive objects in the image, and outputs the sensitive object recognition result.

[0193] In this way, it is possible to determine whether to start the corresponding desensitized object recognition model for processing based on the value of the detection parameter. In other words, only the desensitized object recognition model corresponding to the object type that the user is interested in is started, rather than starting all the desensitized object recognition models in the preset reasoning service module. This can reduce the requirements for computing resources and improve the detection efficiency of sensitive objects; and it provides users with the option of starting the desensitized object recognition model, thereby improving the user's control over the preset reasoning service module. The model version of each desensitized object recognition model can be the version number of the desensitized object recognition model, wherein each version number corresponds one-to-one to each desensitized object recognition model. The reasoning service module retrieves the desensitized object recognition model corresponding to each version number for processing according to each version number.

[0194] The confidence threshold of each desensitized object recognition model is used as a judgment condition for whether each desensitized object recognition model detects the corresponding sensitive object. Specifically, after each desensitized object recognition model completes the recognition processing of the image, it outputs the confidence value of the sensitive object under each object type.

[0195] The desensitized object recognition model compares the confidence value of the sensitive object output by it with the confidence threshold corresponding to each desensitized object recognition model. If the confidence value of the output sensitive object is greater than or equal to its corresponding confidence threshold, it indicates that the desensitized object recognition model has recognized the sensitive object; conversely, if the confidence value of the output sensitive object is less than its corresponding confidence threshold, it indicates that the desensitized object recognition model has not recognized the sensitive object.

[0196] For example, the confidence threshold of the pedestrian object included in the structured data is 0.60. If the confidence value output by the desensitized object recognition model corresponding to the pedestrian object type is 0.65, it indicates that the desensitized object recognition model corresponding to the pedestrian object type has detected the pedestrian object in the image; if the confidence value output by the desensitized object recognition model corresponding to the pedestrian object type is 0.55, it indicates that the desensitized object recognition model corresponding to the pedestrian object type has not detected the pedestrian object in the image.

[0197] Each desensitized object recognition model detects each area in the input image, thereby detecting sensitive objects of each object type, and outputs the confidence value of the sensitive object of each object type; wherein the confidence value of the sensitive object of each object type represents the existence of the object type in the image and the possibility that the object type is a sensitive object; the higher the confidence value, the higher the possibility that the object type in the image is a sensitive object, and vice versa, the lower the possibility that the object type in the image is a sensitive object. The re-inspection mark of each desensitized object recognition model represents whether to perform secondary detection on sensitive objects under certain object types.

[0198] The batch size of each desensitized object recognition model represents the number of images input by the user.

[0199] In one example, the structured data includes a recheck mark of an image region identified by the object recognition model.

[0200] The target type objects in the image are identified by the target recognition model in the preset recognition model framework to obtain sensitive objects in the image, including:

[0201] Feed the image into the object recognition model.

[0202] In response to the re-inspection identification indicating that re-inspection is required, sensitive object areas in the image are detected based on the target recognition model, and second-order detection processing is performed on each sensitive object area to obtain the sensitive objects in the image output by the target recognition model; wherein the sensitive object area is the area in the image where sensitive objects of the target type are located.

[0203] In response to the re-inspection identification indicating that re-inspection is not required, a sensitive object area in the obtained image is determined based on the target recognition model, which is a sensitive object in the image output by the target recognition model.

[0204] Exemplarily, for some specific types of sensitive objects, such as license plate objects, face objects, and sensitive geographic area objects, the desensitized object recognition models corresponding to the above-mentioned sensitive objects are respectively the first desensitized object recognition model, the second desensitized object recognition model, and the third desensitized object recognition model as examples for explanation.

[0205] If the startup parameters indicate that the second-order detection needs to be started, the first desensitized object recognition model first detects the image area where the vehicle exists in the image. Next, the first desensitized object recognition model detects the license plate object in the image area where the vehicle exists and outputs the confidence value corresponding to the license plate object.

[0206] For the license plate object, if the second-order detection is not started, the first desensitized object recognition model will directly detect the license plate object in the entire image. However, since the image also includes other information, for example, the other information may be the house number in the building, and since there is a certain similarity between the house number and the license plate, it is easy for the first desensitized object recognition model to mistakenly recognize the license plate as the license plate object, which will cause an increase in the false detection rate. If the second-order detection is started, the first desensitized object recognition model will only detect the license plate object in the image area where the vehicle exists in the image, thereby avoiding confusion between the house number and the license plate, thereby improving the recognition accuracy of the desensitized object recognition model.

[0207] If the startup parameters indicate that the second-order detection needs to be started, the second desensitized object recognition model first detects the image area where pedestrians exist in the image. Next, the second desensitized object recognition model detects facial objects in the image area where pedestrians exist, and outputs the confidence value corresponding to the facial object.

[0208] For facial objects, if the second-order detection is not started, the second desensitized object recognition model will directly detect facial objects in the entire image. However, since the image also includes other information, such as the portrait on the billboard beside the road, and since there is a certain degree of similarity between the portrait on the billboard and the face, it is easy for the second desensitized object recognition model to mistakenly recognize the portrait on the billboard as a facial object, which will cause an increase in the false detection rate. If the second-order detection is started, the second desensitized object recognition model will only detect facial objects in the image area where there are pedestrians in the image, thereby avoiding confusion between the portrait on the billboard and the face, thereby improving the recognition accuracy of the desensitized object recognition model.

[0209] If the startup parameters indicate that the second-order detection needs to be started, the third desensitized object recognition model first detects the image area where the sensitive geographical area exists in the image. Next, the third desensitized object recognition model detects sensitive geographical area objects in the image area where the sensitive geographical area exists, for example, sensitive geographical area identification plates, and outputs the confidence value corresponding to the license plate object.

[0210] For sensitive geographic area objects, such as sensitive geographic area signs, if the second-order detection is not started, the third desensitized object recognition model will directly detect sensitive geographic area objects in the entire image. However, since the image also includes other information, such as the text on the billboard beside the road, and since there is a certain degree of similarity between the text on the billboard and the text included in the sensitive geographic area sign, it is easy for the third desensitized object recognition model to mistakenly recognize the text on the billboard as a sensitive geographic area sign, which will cause an increase in the false detection rate. If the second-order detection is started, the third desensitized object recognition model will only detect sensitive geographic area objects in the image area where sensitive geographic areas exist in the image, thereby avoiding confusion between the text on the billboard and the sensitive geographic area objects, thereby improving the recognition accuracy of the desensitized object recognition model.

[0211] The following is an example of structured data:

[0212] {

[0213] "classType":[0,1,1,0,1,0,1],

[0214] "modelVersion":[v1.1,v2.1,v3.1,v1.0,v2.0,v2.2,v1.3]

[0215] "confidenceList":[0.50,0.60,0.55,0.50,0.50,0.60,0.70],

[0216] "isSecondDetect":[True,False,False,True,False,False,True],

[0217] "batchSize": 1,

[0218] }

[0219] The content in "classType" represents the object type, which includes various detection parameters. The value of each detection parameter indicates whether to detect sensitive objects of the corresponding object type. For example, if the value of the detection parameter is 1, it means that the object type corresponding to the detection parameter needs to be detected; if the value of the detection parameter is 0, it means that the object type corresponding to the detection parameter does not need to be detected.

[0220] The content in "modelVersion" represents the model version of each desensitized object recognition model. For example, the version numbers of each desensitized object recognition model are v1.1, v2.1, v3.1, v1.0, v2.0, v2.2, and v1.3.

[0221] The content in "confidenceList" represents the confidence threshold, wherein each confidence threshold corresponds to an object type one by one.

[0222] The content in "isSecondDetect" represents the re-check flag, which includes the values ​​of various startup parameters. For example, if the value of the startup parameter is True, it indicates that the desensitized object recognition model corresponding to certain specific object types needs to start the second-order detection; if the value of the startup parameter is False, it indicates that the desensitized object recognition model corresponding to certain specific object types does not need to start the second-order detection.

[0223] The content in "batchSize" represents the value of the batch size. For example, if the value of the batch size is 1, it means that the image input by the user includes one picture.

[0224] S603: Based on a preset load balancing strategy, a preset recognition model framework is selected from at least two initial recognition model frameworks, each of which includes at least one sensitive object recognition model.

[0225] Exemplarily, since a plurality of initial recognition model frameworks are pre-deployed in a computer device, each of which includes at least one sensitive object recognition model, it is necessary to select a preset recognition model framework from the plurality of initial recognition model frameworks.

[0226] A load balancing strategy is pre-stored in the load balancing module of the computer device. The computer device selects a preset recognition model framework from a plurality of initial recognition model frameworks based on the load balancing strategy.

[0227] For example, the load balancing strategy may be a polling method, or a random selection method, or a method of selecting a module that occupies the least amount of computation.

[0228] The polling method is started in the order in which the inference service modules are deployed. For example, see Figure 5 , when the first reasoning service module is in an occupied state, the second reasoning service module will be started.

[0229] The random selection method is to randomly select one of the multiple reasoning service modules and start it. For example, when the first reasoning service module is in an occupied state, the second reasoning service module can be started, or the third reasoning service module can be started, and the result of the random selection shall prevail, regardless of the current occupancy status of each reasoning service module.

[0230] The method of selecting the module with the smallest computational load, also known as the least connection method, is to determine the inference service module with the smallest occupancy based on the current occupancy of each inference service module, and then start it. For example, if the GPU occupancy of the first inference service module is 40%, the GPU occupancy of the second inference service module is 70%, and the GPU occupancy of the first inference service module is 50%, then the new http request will be assigned to the first inference service module with the smallest computational load.

[0231] S604. Determine the sensitive object recognition model corresponding to the object type indicated by the structured data in the preset recognition model framework as the target recognition model; wherein the structured data includes the object type corresponding to the target recognition model.

[0232] Exemplarily, after step S603, a preset recognition model framework is determined, wherein the initial recognition model framework includes at least one sensitive object recognition model. Then, the computer device parses the structured data, obtains the object type corresponding to the target recognition model, and starts the target recognition model corresponding to each object type based on the determined preset recognition model framework according to the object type indicated in the structured data.

[0233] S605. Input the image and the structured data indicated by the image into a preset recognition model framework, and identify the target type object in the image through the target recognition model in the preset recognition model framework to obtain the sensitive object in the image; wherein, the preset recognition model framework includes sensitive object recognition models corresponding to different object types, the target recognition model is a sensitive object recognition model matched with the structured data indicated by the image, and the target type is the object type corresponding to the target recognition model.

[0234] Exemplarily, a computer device inputs an image and structured data into a preset recognition model framework, and calls each desensitized object recognition model under the object type indicated in the structured data to recognize the image, thereby identifying each sensitive object in the image, and outputting a sensitive object recognition result; wherein the sensitive object recognition result includes the location information of each sensitive object.

[0235] For example, the location information can be reflected in the form of coordinates. After the desensitized object recognition model detects the sensitive object, the desensitized object recognition model calibrates the position of the sensitive object in the image in the form of a prediction box. For a prediction box, it has four vertices. Therefore, the position of the sensitive object in the image can be determined by recording the coordinates of the four vertices of the prediction box; for example, the upper left corner coordinates (X1, Y1), the upper right corner coordinates (X2, Y2), the lower right corner coordinates (X3, Y3), and the lower left corner coordinates (X4, Y4) of the prediction box.

[0236] After determining the position of the sensitive object in the image, the computer device performs desensitization processing on each sensitive object in the image according to the position of each sensitive object in the image, thereby outputting a desensitized image.

[0237] The following is an example of the sensitive object recognition results of an image: [

[0239] {“class”:0,

[0240] "coordinates":[901,1004,1035,1004,901,1224,1035,1224],

[0241] "confidence":0.994708776473999,

[0242] "modelVersion": "v1"

[0243] },

[0244] {“class”:1,

[0245] "coordinates":[801,9004,9035,9004,801,1124,9035,1124],

[0246] "confidence":0.947480774637999,

[0247] "modelVersion": "v2"

[0248] },

[0249] … ]

[0251] The values ​​included in "class" correspond to various object types, and this correspondence is pre-set. For example, 0 corresponds to a pedestrian object type, 1 corresponds to a vehicle object type, and 2 corresponds to a sensitive geographic area object type.

[0252] If the desensitized object recognition model detects the corresponding sensitive object, the detection result will include the "class" value of the object type corresponding to the sensitive object; for example, in the above detection result example, it is characterized that the desensitized object recognition model detected pedestrian objects and vehicle objects in the image, but did not detect sensitive geographic area objects.

[0253] The content included in "coordinates" represents the coordinates of the sensitive object in the image. Taking the coordinates of the upper left corner of the image as (0, 0) as an example, in the above detection result sample, the coordinates of the predicted box of the pedestrian object in the image are: upper left corner coordinates (901, 1004), upper right corner coordinates (1035, 1004), lower left corner coordinates (901, 1224), and lower left corner coordinates (1035, 1224); the coordinates of the predicted box of the vehicle object in the image are: upper left corner coordinates (801, 9004), upper right corner coordinates (9035, 9004), lower left corner coordinates (801, 1124), and lower left corner coordinates (9035, 1124).

[0254] The content included in "confidence" represents the confidence value corresponding to the sensitive object output by the desensitized object recognition model. For example, in the above detection result sample, the confidence value of the pedestrian object is 0.994708776473999, and the confidence value of the vehicle object is 0.947480774637999.

[0255] The content included in "modelVersion" represents the version information of the desensitized object recognition model corresponding to the sensitive object. For example, in the above detection result example, the desensitized object recognition model corresponding to the pedestrian object uses version v1, and the desensitized object recognition model corresponding to the vehicle object uses version v2.

[0256] S606: Perform a desensitization operation on the image according to the sensitive object.

[0257] In one example, step S606 may be implemented by the following process: sending the sensitive object to the desensitizing end, so that the desensitizing end performs a desensitizing operation on the image according to the sensitive object.

[0258] Exemplarily, the computer device sends the image and the sensitive object recognition result to the desensitization end, and the desensitization end performs desensitization processing on the image according to the object type of each sensitive object indicated in the sensitive object recognition result and the location information of each sensitive object, thereby obtaining a desensitized image. The specific process of the desensitization processing can be referred to the content in step S601, which will not be repeated here.

[0259] Figure 8 A schematic diagram of the structure of a computer device provided in an embodiment of the present application Figure 1 ,like Figure 8 As shown, a preset recognition model framework is deployed in the computer device, wherein the preset recognition model framework includes sensitive object recognition models corresponding to different object types.

[0260] like Figure 8 As shown, the preset recognition model framework includes a first sensitive object recognition model, a second sensitive object recognition model, a third sensitive object recognition model, ..., the i-th sensitive object recognition model, ..., the N-th sensitive object recognition model. N is a positive integer greater than 1; i is a positive integer greater than or equal to, less than or equal to N.

[0261] The computer device may be, for example, any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., without limitation.

[0262] For example, each sensitive object recognition model can be deployed in a computer device containing a graphics processing unit (GPU).

[0263] It is worth mentioning that Figure 8 The number of sensitive object recognition models shown in does not represent a limitation on the number of sensitive object recognition models.

[0264] Each sensitive object recognition model may correspond to one or more model instances. Each model instance includes the parameters and structure of the corresponding sensitive object recognition model. In other words, each model instance can complete the recognition and processing process of the corresponding sensitive object recognition model.

[0265] In the process of detecting the image, for each sensitive object recognition model, a model instance can be started to detect sensitive objects in the image; or multiple model instances can be started, for example, the first model instance and the second model instance of the first sensitive object recognition model are started to detect sensitive objects in the image in parallel to meet higher desensitization processing requirements.

[0266] like Figure 8 As shown, the computer device acquires an image and receives structured data input by a user; or, the computer device acquires an image and receives structured data sent by a user. Then, the computer device implements a preprocessing operation based on an image preprocessing module in a software development kit. Based on an inference request module in the software development kit, an HTTP request is generated according to the structured data, or, an HTTP request is generated according to the image and the structured data.

[0267] Then, the computer device uses the computing resources of the GPU and the inference service module to retrieve the target recognition model according to the structured data to identify the image and obtain the sensitive objects in the image.

[0268] Then, the computer device obtains the sensitive object in the image based on the reasoning response module, and desensitizes the image based on the desensitization processing operation in the desensitization processing module. The computer device outputs the desensitized image.

[0269] In this embodiment, based on the above embodiment, by enabling the sensitive object recognition model corresponding to the type of sensitive object that the user is interested in according to the type of sensitive object indicated in the structured data, the image is processed to obtain each sensitive object in the image, and then desensitization processing can be performed on each identified sensitive object. In this way, there is no need to enable all sensitive object recognition models in the preset recognition model framework to recognize and process the image, thereby improving the processing efficiency of the system and providing users with greater autonomy in controlling computer devices, so that they can freely select and process the type of sensitive objects of interest.

[0270] In addition, by installing multiple desensitized object recognition models in the inference service module, each desensitized object recognition model specifically detects sensitive objects of a certain object type and outputs the confidence value of each sensitive object; and confidence thresholds for different sensitive objects are set in the structured data. Only when the confidence value of a sensitive object is greater than or equal to the corresponding confidence threshold of the sensitive object, the desensitized object recognition model determines that the sensitive object is detected. In this way, by setting different confidence thresholds according to the characteristics of different object types, a balance can be achieved between the recall rate and precision rate of the desensitized object recognition model, thereby more accurately detecting various sensitive objects in the image and improving the performance of the desensitized object recognition model.

[0271] Furthermore, the second-order detection in the desensitized object recognition model is initiated through the re-inspection mark in the structured data. The desensitized object recognition model first detects the image area where the same object type exists in the image, and then detects sensitive objects in the image area. Compared with directly detecting sensitive objects from all areas in the image, the second-order detection can reduce the probability of confusion between sensitive objects and other objects in the image, thereby reducing the false detection rate of the desensitized object recognition model and improving the detection accuracy of the desensitized object recognition model.

[0272] Fig. 9 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 3 ,like Fig. 9 As shown, the method includes:

[0273] S901, receiving an image and a sensitive object in the image sent by the server; the sensitive object is executed by the server Figure 5 or Figure 6 obtained by the embodiment of the invention.

[0274] Exemplarily, the execution subject of this embodiment is a desensitized terminal, wherein the desensitized terminal can be any of the following: a terminal device, an independent server, a server cluster, a cloud, the Internet of Things, the Internet of Vehicles, a distributed system, etc., without limitation.

[0275] based on Figure 5 or Figure 6 The computer device used in the embodiment of the present invention may be referred to as a "server". The server is based on Figure 5 or Figure 6 After obtaining the sensitive objects in the image, the server sends the image and the sensitive objects in the image to the desensitizing end.

[0276] S902: Perform a desensitization operation on the image according to the sensitive object.

[0277] Exemplarily, the desensitization end performs a desensitization operation on the image according to the sensitive objects in the image to obtain a desensitized image, wherein the sensitive object recognition result includes each sensitive object in the image and the location information of each sensitive object.

[0278] Fig.10 A schematic diagram of the structure of a computer device provided in an embodiment of the present application Figure 2 ,like Fig.10 As shown, a preset recognition model framework is deployed in the computer device, wherein the preset recognition model framework includes sensitive object recognition models corresponding to different object types.

[0279] like Fig.10 As shown, the preset recognition model framework includes a first sensitive object recognition model, a second sensitive object recognition model, a third sensitive object recognition model, ..., the i-th sensitive object recognition model, ..., the N-th sensitive object recognition model. N is a positive integer greater than 1; i is a positive integer greater than or equal to, less than or equal to N.

[0280] Among them, the computer device at this time is the "server".

[0281] Each sensitive object recognition model may correspond to one or more model instances. Each model instance includes the parameters and structure of the corresponding sensitive object recognition model. In other words, each model instance can complete the recognition and processing process of the corresponding sensitive object recognition model.

[0282] In the process of detecting the image, for each sensitive object recognition model, a model instance can be started to detect sensitive objects in the image; or multiple model instances can be started, for example, the first model instance and the second model instance of the first sensitive object recognition model are started to detect sensitive objects in the image in parallel to meet higher desensitization processing requirements.

[0283] like Fig.10 As shown, the computer device (i.e., the "server") obtains an image and receives structured data input by a user; or, the computer device (i.e., the "server") obtains an image and receives structured data sent by a user. Then, the computer device (i.e., the "server") implements a preprocessing operation based on an image preprocessing module in a software development kit. Based on the inference request module in the software development kit, an http request is generated according to the structured data, or an http request is generated according to the image and the structured data.

[0284] Then, the computer device (i.e., the "server") uses the computing resources of the GPU and the inference service module to retrieve the target recognition model according to the structured data to identify the image and obtain the sensitive objects in the image.

[0285] Then, the computer device (i.e., the "server") obtains the sensitive objects in the image based on the reasoning response module, and the computer device (i.e., the "server") sends the image and the sensitive objects in the image to the desensitizing end based on the reasoning response module.

[0286] A desensitization processing module is deployed in the desensitization end, and the desensitization processing module includes a desensitization processing operation. Then, the desensitization end desensitizes the image based on the desensitization processing operation in the desensitization processing module, and the desensitization end outputs the desensitized image.

[0287] The technical effects of this embodiment are similar to those of the previous embodiments, and will not be described in detail here.

[0288] Fig.11 A schematic diagram of a method for desensitizing an image provided in an embodiment of the present application Figure 4 ,like Fig.11 As shown, the method includes:

[0289] S1101: Determine an image to be desensitized and structured data indicated for the image.

[0290] Exemplarily, the user terminal determines the image and structured data to be desensitized. For example, the user terminal receives the image and structured data to be desensitized input by the user; or the user terminal receives the image and structured data to be desensitized sent by other devices.

[0291] In one example, the structured data includes a service address of the target recognition model, an object type corresponding to the target recognition model, a model version of the target recognition model, a confidence threshold of the target recognition model, a review mark of the image area recognized by the target recognition model, a batch size of the target recognition model, and at least one information in the preprocessed image.

[0292] Exemplarily, the various contents included in the structured data can refer to the contents in step S602, which will not be repeated here.

[0293] S1102: Send the image and structured data indicated by the image to the server, so that the server can obtain the sensitive object in the image through the method provided by the above method embodiment.

[0294] Exemplarily, the user terminal sends the image and structured data to be desensitized to the server terminal, and the server terminal identifies and obtains the sensitive object recognition result according to each desensitized object recognition model provided in the above method embodiment. Among them, the sensitive object recognition result includes each sensitive object in the image and the position of each sensitive object. It is worth noting that the sensitive object can be sent by the server terminal or the desensitization terminal, which is not limited here.

[0295] In one example, the step S1102 of “sending the image and structured data indicated by the image to the server” may include the following process:

[0296] Encapsulate the image and structured data indicated by the image into a first communication protocol message, and send the first communication protocol message to the server; or encapsulate the structured data indicated by the image into a second communication protocol message, and send the image and the second communication protocol message indicated by the image to the server.

[0297] Exemplarily, in one case, the user end encapsulates the image and structured data indicated by the image into a first communication protocol message, and sends the first communication protocol message to the server end.

[0298] In another case, the user end encapsulates the structured data indicated by the image into a second communication protocol message, and sends the image and the second communication protocol message for the image to the server end. It should be understood that the user end, the server end, and the desensitizing end can be deployed in the same device (that is, the user end, the server end, and the desensitizing end are the same end), or they can be deployed in different devices (that is, the user end, the server end, and the desensitizing end are different ends); in addition, the server end and the desensitizing end can be deployed in the same device or in different devices.

[0299] In one example, the first communication protocol message is determined based on a first information indication operation triggered by a human-computer interaction interface; and / or the second communication protocol message is determined based on a second information indication operation triggered by the human-computer interaction interface.

[0300] Exemplarily, the manner of determining the first communication protocol message and the second communication protocol message may refer to the content in step S602, which will not be repeated here.

[0301] S1103, receiving sensitive objects in the image returned by the server, and performing a desensitization operation on the image according to the sensitive objects; or, receiving a desensitized image corresponding to the image, where the desensitized image is obtained by performing a desensitization operation on the image according to the sensitive objects.

[0302] Exemplarily, a user end receives an image and a sensitive object in the image sent by a server end; the user end desensitizes the image according to the sensitive object in the image to obtain a desensitized image; and the user end outputs the desensitized image.

[0303] Alternatively, the server performs desensitization processing on the image according to the sensitive objects in the image to obtain a desensitized image; the server sends the desensitized image to the user end; and the user end receives the desensitized image sent by the server.

[0304] Alternatively, the desensitizing end receives the image and the sensitive object in the image sent by the server; the desensitizing end desensitizes the image according to the sensitive object in the image to obtain a desensitized image; the desensitizing end sends the desensitized image to the user end; and the user end receives the desensitized image sent by the desensitizing end.

[0305] In one example, the "desensitizing the image according to the sensitive object" in step S1103 may include the following process:

[0306] At least one of marking, blurring, encryption and occlusion processing is performed on sensitive objects in the image to obtain a desensitized image.

[0307] Exemplarily, the specific content of the desensitization processing operation can refer to the content of step S601 in the previous embodiment, which will not be repeated here.

[0308] The technical effects of this embodiment are similar to those of the previous embodiments, and will not be described in detail here.

[0309] Fig.12 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 1 ,like Fig.12 As shown, the device comprises:

[0310] The determination module 1201 is used to determine the image to be desensitized and the structured data indicated for the image.

[0311] The recognition module 1202 is used to input the image and the structured data indicated by the image into a preset recognition model framework, and identify the target type object in the image through the target recognition model in the preset recognition model framework to obtain the sensitive object in the image; wherein the preset recognition model framework includes sensitive object recognition models corresponding to different object types, the target recognition model is a sensitive object recognition model matched with the structured data indicated by the image, and the target type is the object type corresponding to the target recognition model.

[0312] The desensitization module 1203 is used to perform a desensitization operation on the image according to the sensitive object.

[0313] In one example, the determination module 1201 is specifically configured to: parse the first communication protocol message to obtain an image and structured data indicated by the image; or

[0314] An image is received, and a second communication protocol message for the image is parsed to obtain structured data for the image.

[0315] In one example, the desensitization module 1203 is specifically used to send the sensitive object to the desensitization end, so that the desensitization end performs a desensitization processing operation on the image according to the sensitive object.

[0316] In one example, the recognition module 1202 is specifically used to: structured data includes a confidence threshold for indicating a target recognition model, input an image into the target recognition model, obtain each candidate sensitive object belonging to the target type in the image and the confidence of each candidate sensitive object; determine the candidate sensitive object whose confidence is greater than or equal to the confidence threshold as a sensitive object in the image output by the target recognition model; and / or

[0317] The structured data includes a re-inspection mark of an image area identified by a target recognition model, and the image is input into the target recognition model; in response to the re-inspection mark indicating that re-inspection is required, sensitive object areas in the image are detected based on the target recognition model, and second-order detection processing is performed on each sensitive object area to obtain sensitive objects in the image output by the target recognition model; in response to the re-inspection mark indicating that re-inspection is not required, the sensitive object areas in the obtained image are determined based on the target recognition model as the sensitive objects in the image output by the target recognition model; wherein the sensitive object area is the area in the image where sensitive objects of the target type are located.

[0318] The technical effects of this embodiment are similar to those of the previous embodiments, and will not be described in detail here.

[0319] Fig.13 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 2 ,like Fig.13 As shown, the device comprises:

[0320] The receiving module 1301 is used to receive the image and the sensitive object in the image sent by the server; the sensitive object is the server executing Figure 5 or Figure 6 The embodiment shown is obtained.

[0321] The desensitization module 1302 is used to perform a desensitization operation on the image according to the sensitive object.

[0322] The technical effects of this embodiment are similar to those of the previous embodiments, and will not be described in detail here.

[0323] Fig.14 A schematic diagram of the structure of the image desensitization processing device provided in the embodiment of the present application Figure 3 ,like Fig.14 As shown, the device comprises:

[0324] The determination module 1401 is used to determine the image to be desensitized and the structured data indicated for the image.

[0325] The sending module 1402 is used to send the image and the structured data indicated by the image to the server, so that the server can Figure 5 or Figure 6 The illustrated embodiment obtains sensitive objects in an image.

[0326] Receive sensitive objects in the image returned by the server, and perform a desensitization operation on the image according to the sensitive objects; or, receive a desensitized image corresponding to the image, where the desensitized image is obtained by performing a desensitization operation on the image according to the sensitive objects.

[0327] In one example, the sending module 1402 is specifically configured to:

[0328] encapsulating the image and the structured data indicated by the image into a first communication protocol message, and sending the first communication protocol message to the server; or

[0329] The structured data indicated for the image is encapsulated into a second communication protocol message, and the image and the second communication protocol message for the image are sent to the server.

[0330] In one example, the first communication protocol message is determined based on a first information indication operation triggered through a human-computer interaction interface; and / or

[0331] The second communication protocol message is determined based on a second information indication operation triggered through the human-computer interaction interface.

[0332] In one example, the structured data includes a service address of the target recognition model, an object type corresponding to the target recognition model, a model version of the target recognition model, a confidence threshold of the target recognition model, a review mark of the image area recognized by the target recognition model, a batch size of the target recognition model, and at least one information in the preprocessed image.

[0333] The technical effects of this embodiment are similar to those of the previous embodiments, and will not be described in detail here.

[0334] Fig.15 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. Fig.15 As shown, the computer device 150 provided in this embodiment includes: at least one processor 1501 and a memory 1502. Optionally, the image desensitization processing device 150 also includes a communication component 1503. The processor 1501, the memory 1502 and the communication component 1503 are connected via a bus 1504.

[0335] In a specific implementation process, at least one processor 1501 executes the computer-executable instructions stored in the memory 1502, so that at least one processor 1501 executes the above method.

[0336] The specific implementation process of the processor 1501 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0337] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0338] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-volatile Memory, NVM), such as at least one disk memory.

[0339] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0340] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0341] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0342] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0343] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0344] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0345] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0346] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0347] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0348] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0349] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary technical means in the art not disclosed by the present invention, are not limited to the precise structure described above and shown in the drawings, and may be modified and changed in various ways without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for desensitizing an image, characterized in that: include: Determining an image to be desensitized and structured data indicated for the image; The image and the structured data indicated by the image are input into a preset recognition model framework, and the object of the target type in the image is recognized by the target recognition model in the preset recognition model framework to obtain the sensitive object in the image; wherein the preset recognition model framework includes sensitive object recognition models corresponding to different object types, the target recognition model is a sensitive object recognition model matched with the structured data indicated by the image, and the target type is the object type corresponding to the target recognition model; A desensitization operation is performed on the image according to the sensitive object.

2. The method according to claim 1, characterized in that The determining of the image to be desensitized and the structured data indicated for the image includes: Parsing the first communication protocol message to obtain the image and structured data indicated for the image; or An image is received, and a second communication protocol message for the image is parsed to obtain structured data for the image.

3. The method according to claim 2, characterized in that The first communication protocol message is sent by the client, or the first communication protocol message is determined based on a first information indication operation triggered by a human-computer interaction interface; and / or The second communication protocol message is sent by the client, or the second communication protocol message is determined based on a second information indication operation triggered through a human-computer interaction interface.

4. The method according to claim 1, characterized in that: The desensitizing operation of the image according to the sensitive object includes: The sensitive object is sent to the desensitizing end so that the desensitizing end performs a desensitizing operation on the image according to the sensitive object.

5. The method according to claim 1, characterized in that: The structured data includes a method for indicating an object type corresponding to the target recognition model; before identifying an object of the target type in the image by the target recognition model in the preset recognition model framework to obtain a sensitive object in the image, the method further includes: The sensitive object recognition model corresponding to the object type indicated by the structured data in the preset recognition model framework is determined as the target recognition model.

6. The method according to claim 1, characterized in that The identifying the target type object in the image by the target recognition model in the preset recognition model framework to obtain the sensitive object in the image includes: The structured data includes a confidence threshold for indicating a target recognition model, the image is input into the target recognition model, and each candidate sensitive object belonging to the target type in the image and the confidence of each candidate sensitive object are obtained; the candidate sensitive object whose confidence is greater than or equal to the confidence threshold is determined as the sensitive object in the image output by the target recognition model; and / or The structured data includes a re-inspection mark of the image area identified by the target recognition model, and the image is input into the target recognition model; in response to the re-inspection mark indicating that re-inspection is required, sensitive object areas in the image are detected based on the target recognition model, and second-order detection processing is performed on each of the sensitive object areas to obtain sensitive objects in the image output by the target recognition model; in response to the re-inspection mark indicating that re-inspection is not required, the sensitive object areas in the image obtained are determined based on the target recognition model as the sensitive objects in the image output by the target recognition model; wherein the sensitive object areas are areas in the image where sensitive objects belonging to the target type are located.

7. The method according to claim 1, characterized in that Before inputting the image and the structured data indicated for the image into the preset recognition model framework, the method further includes: selecting the preset recognition model framework from at least two initial recognition model frameworks based on a preset load balancing strategy, each of the initial recognition model frameworks including at least one sensitive object recognition model; and / or Before the target type object in the image is identified by the target recognition model in the preset recognition model framework to obtain the sensitive object in the image, it also includes: performing a preprocessing operation corresponding to the target recognition model on the image according to the preprocessing requirements of the target recognition model; the preprocessing operation, the desensitization processing operation, and the request to call the target recognition model are encapsulated in the same software toolkit.

8. A method for desensitizing an image, characterized in that: include: Receiving an image and a sensitive object in the image sent by a server; The sensitive object is obtained by the server executing the method described in any one of claims 1 to 7; A desensitization operation is performed on the image according to the sensitive object.

9. A method for desensitizing an image, characterized in that: include: Determining an image to be desensitized and structured data indicated for the image; Sending the image and structured data indicated by the image to a server, so that the server obtains the sensitive object in the image by the method according to any one of claims 1 to 7; Receiving the sensitive object in the image returned by the server, and performing a desensitization operation on the image according to the sensitive object; Or, a desensitized image corresponding to the image is received, where the desensitized image is obtained by performing a desensitization processing operation on the image according to the sensitive object.

10. The method according to claim 9, characterized in that The sending the image and the structured data indicated by the image to the server includes: encapsulating the image and the structured data indicated for the image into a first communication protocol message, and sending the first communication protocol message to the server; or The structured data indicated for the image is encapsulated into a second communication protocol message, and the image and the second communication protocol message for the image are sent to the server.

11. The method according to any one of claims 2 to 3 or 10, characterized in that: The first communication protocol message is determined based on a first information indication operation triggered by a human-computer interaction interface; and / or The second communication protocol message is determined based on a second information indication operation triggered through a human-computer interaction interface.

12. The method according to any one of claims 1 to 3 or 9 to 11, characterized in that: The structured data includes the service address of the target recognition model, the object type corresponding to the target recognition model, the model version of the target recognition model, the confidence threshold of the target recognition model, the re-inspection mark of the image area recognized by the target recognition model, the batch size of the target recognition model, and at least one information in the preprocessed image.

13. A computer device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1-7 or 8 or 9-12.

14. A computer-readable storage medium / computer program product, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method according to any one of claims 1-7 or 8 or 9-12.

Images