Electronic screen sensitive picture detection and desensitization method and device, and electronic equipment

Abstract

Embodiments of the present disclosure disclose electronic screen sensitive picture detection and desensitization method, device and electronic equipment. A specific embodiment of the method comprises: acquiring a to-be-detected video frame sequence; acquiring a detected video frame sequence; performing the following processing on the to-be-detected video frame sequence: determining a target video frame; performing frame repetition detection processing on the target video frame to obtain frame repetition information; generating a time sequence detection frame sequence; performing sensitive content detection processing on the time sequence detection frame sequence to obtain a sensitive detection information set; generating desensitization action information; performing update processing on the to-be-detected video frame sequence and the detected video frame sequence; in response to determining that the updated to-be-detected video frame sequence is empty, performing desensitization reinstallation processing on the detected video frame sequence to obtain a desensitized video source file. The embodiment relates to the technical field of computers, and reduces the waste of computing resources and the occupation of data storage space.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, specifically to methods, apparatus, and electronic devices for detecting and desensitizing sensitive images on electronic screens. Background Technology

[0002] Sensitive image detection and desensitization is a key technology for real-time compliance control of content displayed on electronic screens.

[0003] With the rapid development of smart cities and public information dissemination systems, electronic screens have become a core carrier of information interaction. To ensure information security and compliance, sensitive image detection and desensitization technologies have emerged. These technologies require the detection of content presented in video footage, accurately identifying and desensitizing (covering, blurring, or mosaicking) content that is illegal or unsuitable for public occasions (e.g., content involving pornography, violence, or politics). However, since the computing devices equipped on electronic screens generally have limited computing power, the detection and desensitization methods have certain low-computing-power requirements. Furthermore, in live broadcasts of some sporting events, the rapid movement of people in the frame can easily lead to missed desensitization attempts, demanding high accuracy in the detection, location, and desensitization of sensitive content.

[0004] Currently, the following technical problems often arise when detecting and desensitizing sensitive information in images: By performing frame-by-frame detection on the directly acquired video content, focusing on sensitive content within each frame, the video stream contains significant temporal redundancy (i.e., highly similar content across multiple consecutive frames). This leads to repeated calculations of the same sensitive scenes, resulting in a substantial waste of computational resources. While using fixed-interval keyframe extraction reduces computational load, it is highly susceptible to missing short-duration sensitive scenes (such as "flickering" violations) due to insufficient sampling frequency, or to malicious circumvention of detection modes due to keyframe selection biases, leading to serious missed detection risks. Furthermore, detecting on a single frame requires subsequent desensitization operations for each frame, generating a large amount of related desensitization action information and consuming significant data storage space.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not form prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure provide methods, apparatus, electronic devices, and computer-readable media for detecting and desensitizing sensitive images on electronic screens to solve one or more of the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a method for detecting and desensitizing sensitive images on an electronic screen. The method includes: in response to receiving a detection signal sent by a preset electronic screen, acquiring a sequence of video frames to be detected; acquiring a sequence of detected video frames; and, based on the detected video frame sequence, performing the following sensitive detection processing on the sequence of video frames to be detected: determining the first video frame to be detected in the sequence of video frames to be detected as a target video frame; performing frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information; in response to determining that the frame duplication information does not meet a preset duplication condition, generating a temporal detection frame sequence based on the frame duplication information, the target video frame, and the sequence of video frames to be detected; and performing path-specific sensitive content detection processing on the temporal detection frame sequence to obtain... A sensitive detection information set is generated; based on the sensitive detection information set, desensitization action information is generated; based on the time-series detection frame sequence, the video frame sequence to be detected and the detected video frame sequence are updated to obtain the updated video frame sequence to be detected and the updated detected video frame sequence; in response to determining that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence; in response to determining that the updated video frame sequence to be detected is empty, based on the obtained desensitization action information, the detected video frame sequence is desensitized and reassembled to obtain the desensitized video source file and the desensitized video source file is sent to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0009] Secondly, some embodiments of this disclosure provide an electronic screen sensitive image detection and desensitization device, the device comprising: a first acquisition unit configured to acquire a video frame sequence to be detected in response to receiving a detection signal sent by a preset electronic screen; a second acquisition unit configured to acquire a detected video frame sequence; and a detection processing unit configured to perform the following sensitive detection processing on the video frame sequence to be detected based on the detected video frame sequence: determining the first video frame to be detected in the video frame sequence to be detected as a target video frame; performing frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information; in response to determining that the frame duplication information does not meet a preset duplication condition, generating a time-series detection frame sequence based on the frame duplication information, the target video frame, and the video frame sequence to be detected; and performing path de-pathing on the time-series detection frame sequence. Sensitive content detection processing is performed to obtain a sensitive detection information set; based on the sensitive detection information set, desensitization action information is generated; based on the time-series detection frame sequence, the video frame sequence to be detected and the already detected video frame sequence are updated to obtain an updated video frame sequence to be detected and an updated video frame sequence to be detected; in response to determining that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated video frame sequence to be detected; a sending unit is configured to, in response to determining that the updated video frame sequence to be detected is empty, perform desensitization reassembly processing on the detected video frame sequence based on the obtained desensitization action information to obtain a desensitized video source file and send the desensitized video source file to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0010] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0011] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0012] The above-described embodiments of this disclosure have the following beneficial effects: the electronic screen sensitive image detection and desensitization methods of some embodiments of this disclosure reduce the abnormal playback of sensitive images. Specifically, the reason for the waste of computing resources and the occupation of data storage space is that: by performing frame-by-frame detection on the directly acquired video content to be played, and the detection process is performed on the sensitive content of a single frame, due to the large amount of time redundancy in the video stream (i.e., the content of multiple consecutive frames is highly similar), it will lead to repeated calculation of the same sensitive scene, which greatly wastes computing resources. If key frames are extracted for detection at fixed time intervals, although the computing load is reduced, it is very easy to miss short-duration sensitive images (such as "flickering" type violations) due to insufficient sampling frequency, or the detection mode may be maliciously circumvented due to the deviation in key frame selection, resulting in a serious risk of missed detection. Moreover, for single-frame detection, subsequent targeted desensitization operations are required for each image, generating a large amount of related desensitization action information, occupying a large amount of data storage space. Based on this, the electronic screen sensitive image detection and desensitization methods of some embodiments of this disclosure firstly, in response to receiving the detection signal sent by the preset electronic screen, acquire the video frame sequence to be detected. Therefore, the corresponding sequence of video frames to be detected for subsequent processing can be obtained. Next, the sequence of detected video frames is acquired. Then, based on the detected video frame sequence, the following sensitive detection processing is performed on the sequence of video frames to be detected: First, the first video frame to be detected in the sequence is identified as the target video frame. Next, based on the detected video frame sequence, frame duplication detection processing is performed on the target video frame to obtain frame duplication information. This allows us to determine whether the current video frame to be detected highly overlaps with previously processed video frames, reducing multiple detections of similar video frames. Then, in response to the determination that the frame duplication information does not meet the preset duplication condition, a temporal detection frame sequence is generated based on the frame duplication information, the target video frame, and the sequence of video frames to be detected. This allows us to generate adjacent video frames within a certain time window as a temporal detection frame sequence when the target video frame differs significantly from previously processed video frames. When there are significant changes in the image, more frames are extracted in a short time, reducing the possibility of missed detections. Finally, path-specific sensitive content detection processing is performed on the temporal detection frame sequence to obtain a sensitive detection information set. This allows us to obtain sensitive information from image semantics, target and text detection in the image. Next, based on the sensitive detection information set, desensitization action information is generated. This yields the desensitization actions to be performed on the target temporal detection frame sequence. Then, based on the temporal detection frame sequence, the video frame sequence to be detected and the already detected video frame sequence are updated to obtain the updated video frame sequence to be detected and the updated already detected video frame sequence.Next, in response to the determination that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence. Then, in response to the determination that the updated video frame sequence to be detected is empty, based on the obtained desensitization action information, the above detected video frame sequence is desensitized and reassembled to obtain a desensitized video source file, which is then sent to a preset electronic screen for output. This allows for the desensitization of the video images on the electronic screen. Furthermore, because the target video frame to be processed is first repeatedly detected along with all previously detected video frames, the waste of computational resources caused by detecting a large number of duplicate images is reduced. Next, by analyzing the differences between the target detection frame and previously detected video frames, the time interval and frequency of the extracted temporal detection frame sequence are determined, rather than using a single fixed value for extraction. This reduces the risk of missing short-duration sensitive frames (such as "flickering" violations) due to insufficient sampling frequency, or being maliciously circumvented by keyframe selection errors, thus minimizing the risk of missed detections. Simultaneously, a desensitization message is generated within a certain consecutive video frame interval (since the target position in most similar frames does not change significantly), reducing the generation of a large amount of related desensitization action information. This, in turn, reduces the waste of computing resources and the occupation of data storage space. Attached Figure Description

[0013] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0014] Figure 1 This is a flowchart of some embodiments of the electronic screen sensitive image detection and desensitization method according to the present disclosure; Figure 2 These are schematic diagrams illustrating the structure of some embodiments of the electronic screen sensitive image detection and desensitization device according to the present disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0015] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0016] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0017] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0018] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0019] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0020] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Figure 1 A flowchart 100 is shown, illustrating some embodiments of the electronic screen sensitive image detection and desensitization method according to the present disclosure. The electronic screen sensitive image detection and desensitization method includes the following steps: Step 101: In response to receiving the detection signal sent by the preset electronic screen, acquire the video frame sequence to be detected.

[0022] In some embodiments, the execution entity (e.g., a computing device) of the electronic screen sensitive image detection and desensitization method can, in response to receiving a detection signal sent by a preset electronic screen, acquire the input video frame sequence to be detected via a wired or wireless connection. The preset electronic screen can be a pre-defined display screen device used to output images, such as an advertising electronic screen in a public place. The detection signal can be a prompt signal issued by the preset electronic screen to the execution entity indicating that detection is required when a video source is input. The video frame sequence to be detected can be a sequence of video frames received by the preset electronic screen that needs to be output as video. Each video frame to be detected has a corresponding timestamp. The execution entity can be a computing device configured on the preset electronic screen.

[0023] It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future wireless connection methods.

[0024] Step 102: Obtain the sequence of detected video frames.

[0025] In some embodiments, the execution entity may acquire a sequence of detected video frames. In practice, the execution entity may acquire the sequence of detected video frames from a preset storage device. The sequence of detected video frames may be a sequence obtained by arranging previously detected video frames in the order they appear in the sequence of video frames to be detected. The preset storage device may include, but is not limited to, at least one of the following: a conventional hard disk drive (HDD), a SATA solid-state drive (SATA SSD), and an NVMe M.2 solid-state drive.

[0026] Step 103: Based on the detected video frame sequence, perform the following sensitive detection processing on the video frame sequence to be detected: Step 1031: The first video frame to be detected in the sequence of video frames to be detected is determined as the target video frame.

[0027] In some embodiments, the execution entity may determine the first video frame to be detected in the above-mentioned video frame sequence as the target video frame.

[0028] Step 1032: Based on the detected video frame sequence, perform frame duplication detection processing on the target video frame to obtain frame duplication information.

[0029] In addressing the technical problems mentioned above, the malicious insertion of flicker-type sensitive information in the intended application scenario often presents the following challenges: When performing extractive detection on video frame sequences, using fixed time intervals to extract keyframes reduces computational load but easily leads to missed short-duration sensitive frames (such as "flickering" violations) due to insufficient sampling frequency, or detection modes being maliciously circumvented due to keyframe selection bias, resulting in serious missed detection risks. This application scenario requires the following characteristics: analyzing the magnitude of differences in frame changes to obtain frame repetition information representing the degree of repetition, and then extracting subsequent video frames based on different rates of frame change. Faced with these technical problems, we decided to adopt the following solution: In some embodiments, the aforementioned execution entity may perform frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information.

[0030] In some optional implementations of certain embodiments, the aforementioned execution entity can perform frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information through the following steps: The first step is to generate a sequence of adjacent video frames based on the detected video frame sequence. In practice, firstly, the executing entity can determine a predetermined number of detected video frames from the reverse order of the detected video frame sequence as a predetermined number of adjacent video frames. Then, the resulting adjacent video frames are arranged in the order they appear in the detected video frame sequence to form the next adjacent video frame sequence. The predetermined number can be a pre-set number of video frames selected from the detected video frame sequence.

[0031] The second step involves performing the following processing on each adjacent video frame in the target video frame and its adjacent video frame sequence: The first sub-step involves generating single-point pixel differences between the target video frame and adjacent video frames to obtain single-point pixel difference information. In practice, firstly, the executing entity can determine the absolute difference between each corresponding pixel in the target video frame and the adjacent video frames as pixel differences, obtaining various pixel differences. Next, the number of pixel differences less than a first preset value is determined as a first value. Then, the total number of pixels in the target video frame is determined as a second value. Finally, the ratio of the first value to the second value is determined as the single-point pixel difference information. The first preset value can be a pre-defined maximum value of two-point pixel differences. The single-point pixel difference information can represent the proportion of pixels with relatively small changes in the image among all pixels.

[0032] The second sub-step involves performing global pixel difference analysis on the target video frame and adjacent video frames to obtain global pixel difference information. In practice, firstly, the execution entity can determine the grayscale histogram of the target video frame as the first grayscale histogram. Next, the grayscale histograms of the adjacent video frames are determined as the second grayscale histogram. Then, the Bach coefficient values ​​of the first and second grayscale histograms are used to determine the global pixel difference information.

[0033] The third sub-step involves generating structural visual differences between the target video frame and adjacent video frames to obtain structural similarity information. In practice, the executing entity can first determine the structural similarity information using the Structure Similarity Index Measure (SSIM) between the target video frame and the adjacent video frames.

[0034] The fourth sub-step generates inter-frame difference information based on single-point pixel difference information, global pixel difference information, and structural similarity information. In practice, the executing entity can determine the inter-frame difference information by summing the values ​​of the single-point pixel difference information, global pixel difference information, and structural similarity information according to a preset weight value. The preset weight value can be a pre-defined proportion of the summation of the values ​​of the single-point pixel difference information, global pixel difference information, and structural similarity information.

[0035] The third step is to generate frame repetition information based on the obtained inter-frame difference information. In practice, the executing entity can determine the inter-frame difference information with the smallest value among the various inter-frame difference information as the frame repetition information.

[0036] The above technical solution, combined with steps 1033-104 and related content, serves as an inventive point of this disclosure, solving the technical problem of "causing serious missed detection risk." Factors leading to serious missed detection risk are often as follows: When performing extractive detection on a video frame sequence, if keyframes are extracted at fixed time intervals for detection, although the computational load is reduced, it is very easy to miss short-duration sensitive scenes (such as "flickering" type violations) due to insufficient sampling frequency, or the detection mode may be maliciously circumvented due to keyframe selection deviation, resulting in serious missed detection risk. Solving the above factors can reduce the missed detection risk. To achieve this effect, firstly, based on the detected video frame sequence, a sequence of adjacent video frames is generated. Thus, the latest detected video frame can be extracted for subsequent comparison with the target video frame. Next, the following processing is performed on each adjacent video frame in the target video frame and adjacent video frame sequences: First, single-point pixel difference generation processing is performed on the target video frame and adjacent video frames to obtain single-point pixel difference information. Therefore, the differences between each pixel in the target video frame and pixels at the same location in adjacent video frames can be compared to obtain single-point pixel difference information. Next, global pixel difference analysis is performed on the target video frame and adjacent video frames to obtain global pixel difference information. This allows us to obtain the differences between the overall grayscale features of the target video frame and those of adjacent video frames. Then, structural visual difference generation processing is performed on the target video frame and adjacent video frames to obtain structural similarity information. This allows us to obtain the differences in visual effects between the target video frame and adjacent video frames, and their similarity in visual structure. Next, based on the single-point pixel difference information, global pixel difference information, and structural similarity information, inter-frame difference information is generated. Finally, based on the obtained inter-frame difference information, frame repetition information is generated. The process begins by extracting a predetermined number of newly detected video frames from the already detected video frame sequence as adjacent video frame sequences to be detected against the target video frame. Next, each adjacent video frame in the adjacent video frame sequence is compared pixel-level with the target video frame: First, each pixel in the adjacent and target video frames is compared to obtain the proportion of pixels with significant changes, which is used as single-point pixel difference information. Then, the Baumel coefficient between the grayscale histograms of the adjacent and target video frames is used as global pixel difference information to represent the overall similarity of the images. Next, a structural similarity information index is used to represent the similarity of the visual structure between the two images. Finally, the obtained single-point pixel difference information, global pixel difference information, and structural similarity information are combined to obtain inter-frame difference information, which comprehensively represents the similarity between an adjacent video frame and the target video frame. Finally, the inter-frame difference information with the smallest value among all the obtained inter-frame difference information is determined as frame repetition information.The smaller the value of the frame repetition information, the less similar the two frames are, and the greater the degree of change in the image. Combined with subsequent steps 1033-104, different temporal extraction information (the extraction frequency of the video frame sequence to be detected) is determined by different frame repetition information. The smaller the value of the frame repetition information, the higher the extraction frequency, and the more video frames to be detected are extracted within a certain time interval, reducing the risk of serious missed detection caused by insufficient sampling frequency and missing sensitive images with short duration.

[0037] Step 1033: In response to determining that the frame repetition information does not meet the preset repetition condition, a temporal detection frame sequence is generated based on the frame repetition information, the target video frame, and the video frame sequence to be detected.

[0038] In some embodiments, the execution entity may, in response to determining that the frame repetition information does not meet a preset repetition condition, generate a temporal detection frame sequence based on the frame repetition information, the target video frame, and the sequence of video frames to be detected. The preset repetition condition may be that the frame repetition information is less than a preset threshold value. The preset threshold value may be a pre-defined minimum value for the frame repetition information. It should be noted that when the frame repetition information is greater than the preset threshold value, it indicates that the target detection frame is extremely similar to the most recently detected frame. The target detection frame is then deleted from the sequence of video frames to be detected and added to the end of the sequence of detected video frames to complete the update of both the sequence of video frames to be detected and the sequence of detected video frames. Based on the updated sequence of detected video frames, the aforementioned sensitive detection processing is then performed on the updated sequence of video frames to be detected.

[0039] In some optional implementations of certain embodiments, the aforementioned execution entity can generate a temporal detection frame sequence based on frame repetition information, the target video frame, and the sequence of video frames to be detected through the following steps: The first step is to obtain the temporal extraction information corresponding to the frame repetition information. This temporal extraction information includes the extracted frame value and the extraction time window information. The extracted frame value can be a pre-defined number of video frames extracted from the time window video frame sequence. The extraction time window information can be a pre-defined time interval corresponding to the timestamps of the extracted video frames from the video frame sequence to be detected. Different values ​​in the frame repetition information will result in different temporal extraction information. For example, when the value corresponding to the frame repetition information is greater than a certain value, it indicates a higher probability of a sudden change in the scene, and the corresponding extracted frame value can be higher (representing more extracted frames) and the extraction time window information can be smaller (a shorter analysis time per session). The time window video frame sequence can be a sequence composed of each video frame to be detected whose timestamp falls within the time range indicated by the extraction time window information, arranged in chronological order within the video frame sequence to be detected. In practice, the executing entity can obtain the temporal extraction information corresponding to the frame repetition information from a pre-set storage device.

[0040] The second step involves extracting time window information and processing the video frame sequence to be detected using time window video frame extraction to obtain a time window video frame sequence. In practice, firstly, the executing entity can identify each video frame in the video frame sequence whose timestamp falls within the extracted time window information as a time window video frame. Then, the obtained time window video frames are arranged according to their chronological order within the video frame sequence to obtain the time window video frame sequence. The third step involves dynamically extracting frames from the time-window video frame sequence based on the extracted frame values ​​to obtain a timing detection frame sequence. In practice, firstly, the execution entity can determine the number of time-window video frames in the time-window video frame sequence as the third value. Next, the ratio of the extracted frame value to the third value is rounded down to obtain the fourth value. Then, starting with the first time-window video frame in the time-window video frame sequence and using the fourth value as an interval, a number of time-window video frames corresponding to the extracted frame values ​​are extracted from the time-window video frame sequence as timing detection frames. Finally, the obtained timing detection frames are arranged according to their chronological order in the time-window video frame sequence to obtain the timing detection frame sequence.

[0041] Step 1034: Perform path-specific sensitive content detection processing on the time-series detection frame sequence to obtain a sensitive detection information set.

[0042] In some embodiments, the aforementioned execution entity may perform path-specific sensitive content detection processing on the time-series detection frame sequence to obtain a sensitive detection information set.

[0043] In some optional implementations of certain embodiments, the aforementioned execution entity can perform path-specific sensitive content detection processing on the temporal detection frame sequence through the following steps to obtain a sensitive detection information set: The first step is to perform the following processing on each timing detection frame in the timing detection frame sequence: The first sub-step involves preprocessing the timing detection frames to obtain preprocessed video frames. In practice, the aforementioned execution entity can convert the timing detection frames into a preset format, obtaining the converted timing detection frames as preprocessed video frames. The preset format can be a pre-defined unified format for subsequent processing of the timing detection frames (e.g., color space conversion to a unified space, size scaling to the same size, normalization, etc.).

[0044] The second sub-step involves performing image semantic category detection on the preprocessed video frames to obtain image semantic category information. In practice, the aforementioned execution entity can input the preprocessed video frames into a pre-trained image classification model. The image semantic category information can be a category name representing the possible types of content present in the preprocessed video frames, along with their corresponding confidence scores. For example, the image semantic category information could be {[Normal: 0.23], [Violence: 0.89], [Pornography: 0.14], [Political: 0.41]}. The pre-trained image classification model can be a neural network model (e.g., a ResNet model) that takes the preprocessed video frames as input and the image semantic category information as output. As an example, at least one sample preprocessed video frame and the corresponding sample image semantic category information for each sample preprocessed video frame can be obtained first. Then, each sample preprocessed video frame is used as input, and the corresponding image semantic category information is used as the desired output to train the image classification model.

[0045] The third sub-step involves inputting the pre-processed video frames into a pre-trained sensitive target detection model to obtain a sensitive target detection information set. This pre-trained sensitive target detection model includes a backbone network, a neck network, and a detection head.

[0046] The fourth sub-step involves performing sensitive text recognition processing on the preprocessed video frames to obtain sensitive text information. In practice, firstly, the aforementioned execution entity can input the preprocessed video frames into the Tesseract OCR tool to obtain image text information. Then, the sensitive text information is matched against a preset sensitive word library to obtain sensitive text information. The sensitive text information can be composed of sensitive words matched during the matching process. It should be noted that if no sensitive words are matched during the matching process, the sensitive text information can be empty. The Tesseract OCR tool is an open-source optical character recognition engine maintained by Google. The image text information can be the text contained in the preprocessed video frames extracted from them. The image text information can include text information and text location information. The text information can be the text contained in the preprocessed video frames extracted from them. The text location information can be information indicating the position of the text information within the preprocessed video frames. The aforementioned location information can be the coordinates of the four vertex corners of the rectangle containing the text. The aforementioned preset sensitive word database can be a pre-defined database containing various sensitive words.

[0047] The fifth sub-step involves determining the image semantic category information, the sensitive target detection information set, and the aforementioned sensitive text information as sensitive detection information.

[0048] The second step is to define the obtained sensitive detection information as a sensitive detection information set. Each sensitive detection information in the set corresponds to one temporal detection frame in the aforementioned temporal detection frame sequence.

[0049] In some optional implementations of certain embodiments, the aforementioned execution entity can input pre-processed video frames into a pre-trained sensitive target detection model through the following steps to obtain a sensitive target detection information set: The first step is to input the preprocessed video frames into the backbone network to obtain the first feature fusion map. The backbone network adopts the CSPDarknet53 structure, which includes: Focus layer: Input 640×640×3 (preprocessed video frames), which reduces the image width and height to half and expands the number of channels to 12 through slicing, resulting in an output size of 320×320×12, used for downsampling and preserving original image information; First convolutional layer (Conv): Input 320×320×12, processed by 3×3 convolution (stride 2), output size 160×160×32; First C3 module: Input 160×160×32, using a cross-stage local network structure, dividing the feature map into two paths, convolving them separately and then fusing them, output size 160×160×64; Second convolutional layer (Conv): Input 160×160×64, downsampled by 3×3 convolution (stride 2), output size 80×80×128. The second C3 module: input 80×80×128, output size 80×80×128, extracts mid-level semantic features; the third convolutional layer (Conv): input 80×80×128, after 3×3 convolution (stride 2) downsampling, output size 40×40×256; the third C3 module: input 40×40×256, output size 40×40×256, extracts high-level semantic features; the fourth convolutional layer (Conv): input 40×40×256, after 3×3 convolution (stride 2) downsampling, output size 20×20×512; the fourth C3 module: input 20×20×512, output size 20×20×512; the SPPF module: input 20×20×512, through multi-scale max pooling to fuse different receptive field features, output size 20×20×512 (first feature fusion map). The first feature fusion image mentioned above can be an image with a size of 20×20×512 output by the SPPF module in the backbone network.

[0050] The second step involves inputting the first feature fusion map into the neck network to perform feature enhancement processing, resulting in the first enhanced feature map. The neck network employs a PANet structure, including a top-down upsampling path and a bottom-up downsampling path. Specifically, it includes: a first upsampling branch: reducing the 20×20×512 feature map (first feature fusion map) output by SPPF to 256 channels via a 1×1 convolution, then performing a 2x upsampling to obtain a 40×40×256 feature map. This 40×40×256 feature map is then concatenated with the 40×40×256 feature map output by the third C3 module in the backbone network. After concatenation, the concatenated feature map is processed by the fifth C3 module, with an output size of 40×40×256. A second upsampling branch: reducing the 40×40×256 feature map to 128 channels via a 1×1 convolution, then performing a 2x upsampling to obtain an 80×80×128 feature map. This 80×80×128 feature map is then concatenated with the 80×80×128 feature map output by the second C3 module in the backbone network. The images are stitched together, and then processed by the sixth C3 module, with an output size of 80×80×128, forming the first detection scale input. The first downsampling branch downsamples the 80×80×128 feature map to 40×40×256 through a 3×3 convolution (stride of 2), and stitches it with the 40×40×256 feature map output from the first upsampling branch. After stitching, it is processed by the seventh C3 module, with an output size of 40×40×256, forming the second detection scale input. The second downsampling branch downsamples the 40×40×256 feature map to 20×20×512 through a 3×3 convolution (stride of 2), and stitches it with the 20×20×512 feature map output from SPPF. After stitching, it is processed by the eighth C3 module, with an output size of 20×20×512 (the first feature enhancement map), forming the third detection scale input. The aforementioned first feature enhancement image can be an image whose features are enhanced after upsampling path processing and downsampling path processing.

[0051] The third step involves inputting the first feature enhancement map into the detection head to perform multi-scale feature detection processing, resulting in a feature detection map set. The detection head comprises three independent detection layers, each corresponding to a different scale output: First detection layer (small target detection): Input 80×80×128, mapped to an 80×80×[3×(4+1+N)] feature map via a 1×1 convolution, where 3 represents the number of anchor boxes per grid point, 4 represents the bounding box coordinates, 1 represents the target confidence score, and N represents the number of tampering categories; Second detection layer (medium target detection): Input 40×40×256, mapped to a 40×40×[3×(4+1+N)] feature map via a 1×1 convolution; Third detection layer (large target detection): Input 20×20×512, mapped to a 20×20×[3×(4+1+N)] feature map (feature detection map) via a 1×1 convolution. Each feature detection map in the above feature detection map set can be a feature map output by the three independent detection layers mentioned above. Each feature detection map in the above feature detection map set can also be a feature map containing the target detected by the detection head.

[0052] The fourth step involves applying non-maximum suppression (NMS) to the feature detection map to obtain a sensitive target detection information set. In practice, the aforementioned execution entity can process the feature detection map using NMS to obtain the sensitive target detection information set. Each sensitive target detection information in this set can include a candidate bounding box, confidence score, and target category. This sensitive target detection information can be obtained by detecting images in preprocessed video frames, obtaining information representing the aforementioned categories (target categories), target locations (candidate bounding boxes), and reliability scores (confidence scores) of sensitive targets within the image.

[0053] Step 1035: Generate desensitization action information based on the sensitive detection information set.

[0054] In some embodiments, the aforementioned executing entity may generate desensitization action information based on the sensitive detection information set.

[0055] In some optional implementations of certain embodiments, the aforementioned execution entity can generate desensitization action information based on the sensitive detection information set through the following steps: The first step is to perform normalization mapping on the sensitive detection information set to obtain a normalized mapping information set. In practice, firstly, the aforementioned executing entity can map the values ​​contained in each of the sensitive detection information to the range [0, 100], and determine the mapped sensitive detection information as the normalized mapping information. Then, each normalized mapping information is determined as a normalized mapping information set. The normalized mapping information can include image semantic category information, sensitive target detection information set, and sensitive text information. The image semantic category information, sensitive target detection information set, and sensitive text information in the above normalized mapping information are all information whose values ​​have been mapped to the range [0, 100].

[0056] The second step involves generating multi-frame mapping information based on the normalized mapping information set. In practice, firstly, the aforementioned execution entity summarizes the semantic category information of each image in the normalized mapping information set to generate multi-frame semantic information. Next, it summarizes and merges the sensitive target detection information sets in the normalized mapping information set (for the same target category, retaining the target with the highest confidence value, retaining all target candidate boxes, and marking the frame number corresponding to each target candidate), resulting in a multi-frame sensitive target information set. Then, it merges the sensitive text information from the normalized mapping information set into a single set, resulting in multi-frame sensitive text information. Finally, the multi-frame semantic information, the multi-frame sensitive target information set, and the multi-frame sensitive text information are defined as the multi-frame mapping information. As an example, the semantic category information of each image in the above normalized mapping information set can be {[Normal: 23], [Violence: 89], [Pornography: 14], [Politicism: 41]}; {[Normal: 20], [Violence: 92], [Pornography: 11], [Politicism: 35]}. The detection information set of each sensitive target in the above normalized mapping information set can be {[Pedestrian 1: Box A, Confidence 72][Vehicle 1: Box B, Confidence 85]} corresponding to frame 1, and {[Pedestrian 1: Box C, Confidence 88][Vehicle 1: Box D, Confidence 79]} corresponding to frame 2. If each sensitive text information in a mapping information set can be {attack} or {beating}, then the resulting multi-frame semantic information can be {[normal: 23], [violence: 89], [pornography: 14], [politics: 41]; [normal: 20], [violence: 92], [pornography: 11], [politics: 35]}. The above multi-frame sensitive target information set can be {[pedestrian 1: frame A-frame 1, frame C-frame 2, confidence 0.88], [vehicle 1: frame B-frame 1, frame D-frame 2, confidence 0.85]}. The above multi-frame sensitive text information can be {attack, beating}.

[0057] Third, in response to the multi-frame mapping information satisfying the preset multi-frame mapping conditions, the following steps are performed on each sensitive detection information in the sensitive detection information set: The first sub-step involves performing target localization processing on the sensitive target detection information set in the sensitive detection information to obtain a sensitive target information set. The aforementioned preset multi-frame mapping condition can be that the numerical value corresponding to the abnormal category in the multi-frame semantic information is greater than the fourth value, or that at least one confidence value in the multi-frame sensitive target information set is greater than the fourth value. In practice, firstly, for each target candidate box in the aforementioned sensitive target detection information set, the following steps are performed: First, the temporal detection frame corresponding to the aforementioned sensitive target detection information set is determined as the target temporal detection frame. Next, the image corresponding to the target candidate box in the aforementioned target temporal detection frame is cropped to obtain the cropped image as the target image. Then, the aforementioned sensitive target detection information and the aforementioned target image are combined to form an information group as sensitive target information. Finally, the obtained sensitive target information sets are determined as the sensitive target information set.

[0058] The second sub-step involves performing text localization processing on the sensitive text information in the sensitive detection information to obtain text localization information. In practice, firstly, the temporal detection frame corresponding to the aforementioned sensitive detection information is determined as the target temporal detection frame. Next, the image corresponding to the aforementioned text localization information in the aforementioned target temporal detection frame is cropped to obtain the cropped image as the target text image. Then, the aforementioned sensitive target detection information and the aforementioned target text image are combined to form an information group, which serves as the text localization information.

[0059] The third sub-step generates single-frame desensitization action information based on image semantic category information, sensitive target information set, and text location information. In practice, firstly, the executing entity can determine the action type by identifying the category with the highest confidence value in the image semantic category information. Then, the sensitive target information set, the text location information, and the action type are used to determine the single-frame desensitization action information. The action types can be normal, slightly blurred, very blurred, or obscured. For example, if the category with the highest confidence value in the image semantic category information is "politics," the corresponding action type could be "slightly blurred"; if the category with the highest confidence value in the image semantic category information is "bloodshed," the corresponding action type could be "obscured."

[0060] The fourth step involves processing the obtained single-frame desensitization action information into desensitization frame intervals to obtain desensitization action information. In practice, the executing entity can use the timestamp corresponding to the first frame in the aforementioned time-series detection frame sequence as the start point and the timestamp corresponding to the last frame as the end point to generate a time interval as the action time interval. Then, single-frame desensitization action information of the same action type is grouped together to obtain each single-frame desensitization action information group. Finally, the aforementioned action time interval, the aforementioned single-frame desensitization action information groups, and the aforementioned multi-frame mapping information are determined as the desensitization action information.

[0061] In addressing the aforementioned technical problems in the process of adopting technical solutions, the application scenario—desensitization processing of sports event broadcast videos—often presents the following technical challenges: In desensitization processing of sports event broadcast videos, targets in some events move too quickly, making it difficult for static desensitization actions on a single frame to match the motion contours, easily leading to the exposure of local sensitive pixels and resulting in missed desensitization in some areas. Considering the following requirements for this application scenario: in scenarios such as sports events where targets move rapidly, the desensitization action should reduce its reliance on single frames. Faced with these technical problems, we decided to adopt the following solution: In some optional implementations of certain embodiments, the aforementioned execution entity may further generate desensitization action information based on the sensitive detection information set through the following steps: The first step is to generate a sensitive information mask set based on the sensitive detection information set. Each sensitive information mask in the above sensitive information mask set corresponds to a temporal detection frame in the above temporal detection frame sequence. In practice, firstly, for each sensitive detection information in the above sensitive detection information set, the execution entity can perform the following processing steps: First, determine the temporal detection frame corresponding to the above sensitive detection information as the temporal detection frame to be processed. Next, determine the sensitive target detection information set in the sensitive detection information as the target sensitive target detection information set. Then, for each target sensitive target detection information in the above target sensitive target detection information set, the execution entity can perform the following steps: First, determine the candidate boxes in the target sensitive target detection information as target candidate boxes. Next, determine the image of the target candidate box in the region corresponding to the temporal detection frame to be processed as a region sub-image. Then, construct a blank binary mask matrix with the same size as the temporal detection frame to be processed. Then, assign the pixel regions corresponding to the region sub-image within the binary mask matrix to valid mask values, while keeping the remaining regions as invalid mask values, and determine the assigned mask matrix as sensitive mask information. Finally, the obtained sensitive information masks are determined as a sensitive information mask set. In this set, the coordinates of each pixel in the sensitive information mask are preserved as the coordinates of its corresponding position in the temporal detection frame. Valid mask values ​​represent coordinates within a sensitive area of ​​the mask matrix, requiring desensitization processing for that pixel location in the original image; these values ​​are typically set to a constant of 1. Invalid mask values ​​represent coordinates within a normal, non-sensitive area of ​​the mask matrix, preserving the original image for that pixel location without desensitization processing; these values ​​are typically set to a constant of 0.

[0062] The second step is to obtain the initial desensitization mask information sequence. In practice, the aforementioned execution entity can obtain the initial desensitization mask information sequence from a preset storage device. This initial desensitization mask information sequence can be used to store the sensitive information mask after certain fusion processing of the sensitive information mask. The initial desensitization mask information sequence can initially be zero.

[0063] The third step involves performing the following mask fusion steps on the temporal detection frame sequence and the sensitive detection information set, based on the desensitization mask information sequence: The first sub-step involves determining the first timing detection frame in the timing detection frame sequence as the first timing detection frame, and determining the second timing detection frame as the second timing detection frame.

[0064] The second sub-step involves adding the sensitive information mask corresponding to the first time-series detection frame to the initial desensitization mask information sequence to update the initial desensitization mask information sequence, and then determining the updated initial desensitization mask information sequence as the desensitization mask information sequence. In practice, the aforementioned executing entity can add the sensitive information mask corresponding to the first time-series detection frame to the end of the initial desensitization mask information sequence to update it.

[0065] The third sub-step involves performing bidirectional optical flow estimation on the first and second temporal detection frames to obtain a forward optical flow vector information set and a reverse optical flow vector information set. In practice, firstly, the execution entity can use a dense optical flow algorithm to calculate the pixel displacement vector from the first temporal detection frame to the second temporal detection frame pixel-by-pixel as the forward optical flow vector information, and determine each obtained forward optical flow vector information set. Then, the dense optical flow algorithm is used to calculate the pixel displacement vector from the second temporal detection frame to the first temporal detection frame pixel-by-pixel as the reverse optical flow vector information, and determine each obtained reverse optical flow vector information set. Each forward optical flow vector in the forward optical flow vector information set corresponds to one pixel in the first temporal detection frame and one pixel in the second temporal detection frame. Similarly, each reverse optical flow vector in the reverse optical flow vector information set corresponds to one pixel in the first temporal detection frame and one pixel in the second temporal detection frame.

[0066] The fourth sub-step involves performing mask deduction and fusion processing on the first and second temporal detection frames based on the desensitization mask information sequence, the forward optical flow vector information set, and the reverse optical flow vector information set to obtain the corrected fusion mask.

[0067] The fifth sub-step involves generating a mask time interval based on the first and second timing detection frames. In practice, the executing entity can determine the time interval between the timestamp corresponding to the second timing detection frame and the timestamp corresponding to the first timing detection frame as the mask time interval.

[0068] The sixth sub-step involves generating desensitization mask information based on the corrected fusion mask and the mask time interval, and adding the desensitization mask information to the desensitization mask information sequence to update the sequence. In practice, the executing entity can determine the corrected fusion mask and the mask time interval as the desensitization mask information, and add the desensitization mask information to the end of the desensitization mask information sequence to update it.

[0069] The seventh sub-step involves deleting the first timing detection frame from the timing detection frame sequence to update the timing detection frame sequence.

[0070] The eighth sub-step involves, in response to the determination that the updated timing detection frame sequence does not meet the preset condition, performing the above mask fusion step again on the updated timing detection frame sequence and the updated sensitive detection information set based on the updated desensitization mask information sequence. The preset condition may be that only one timing detection frame exists in the updated timing detection frame sequence.

[0071] The ninth sub-step involves, in response to the determination that the updated timing detection frame sequence meets preset conditions, performing desensitization image interval processing on the updated desensitization mask information sequence to obtain desensitization action information. In practice, the executing entity can use the timestamp corresponding to the first frame in the timing detection frame sequence as the start point and the timestamp corresponding to the last frame as the end point to generate a time interval as the action time interval. Then, the action time interval and the updated desensitization mask information sequence are determined as the desensitization action information.

[0072] The above technical solution, combined with steps 1035-104 and related content, serves as an inventive point of this disclosure, solving the technical problem of "exposed local sensitive pixels, resulting in partial area desensitization failure." Factors leading to exposed local sensitive pixels and partial area desensitization failure often include: in the desensitization processing of sports broadcast videos, targets in some events move too fast, making it difficult for a single-frame static desensitization action to match the motion contour, easily leading to exposed local sensitive pixels and partial area desensitization failure. Solving these factors can improve the plate recognition effect. To achieve this effect, firstly, a sensitive information mask set is generated based on the sensitive detection information set, where each sensitive information mask in the sensitive information mask set corresponds to a temporal detection frame in the temporal detection frame sequence. Thus, a corresponding masking mask is generated for the region where sensitive targets are detected. Next, an initial desensitization mask information sequence is obtained. Then, based on the initial anonymization mask information sequence, the following mask fusion steps are performed on the temporal detection frame sequence and the sensitive detection information set: First, the first temporal detection frame in the temporal detection frame sequence is determined as the first temporal detection frame, and the second temporal detection frame is determined as the second temporal detection frame. Next, the sensitive information mask corresponding to the first temporal detection frame is added to the initial anonymization mask information sequence to update the initial anonymization mask information sequence, and the updated initial anonymization mask information sequence is determined as the anonymization mask information sequence. Then, bidirectional optical flow estimation processing is performed on the first and second temporal detection frames to obtain the forward optical flow vector information set and the reverse optical flow vector information set. Thus, a corresponding offset vector can be generated for each pixel in the preceding and following images, which is used to compare the corresponding positions of the masks in the preceding and following images for subsequent fusion. Then, based on the anonymization mask information sequence, the forward optical flow vector information set, and the reverse optical flow vector information set, mask deduction fusion processing is performed on the first and second temporal detection frames to obtain the corrected fused mask. Next, a mask time interval is generated based on the first and second time-series detection frames. Then, the corrected fusion mask and the mask time interval are used to determine the desensitization mask information, and this desensitization mask information is added to the desensitization mask information sequence to update the sequence. Then, the first time-series detection frame is deleted from the time-series detection frame sequence to update the sequence. Next, in response to the determination that the updated time-series detection frame sequence does not meet a preset condition, the above mask fusion steps are performed again on the updated time-series detection frame sequence and the updated sensitive detection information set based on the updated desensitization mask information sequence. Thus, mask fusion can be performed cyclically even when there are still more than one time-series detection frame in the time-series detection frame sequence. Finally, in response to the determination that the updated time-series detection frame sequence meets the preset condition, the updated desensitization mask information sequence is subjected to desensitization image interval processing to obtain desensitization action information.This is because the sensitive detection information of the detected sensitive content is first converted into a mask at the corresponding location, so that the mask can be fused between the preceding and following frames. Next, forward optical flow vector information sets and reverse optical flow vector information sets are generated based on the positional changes of corresponding points in the two consecutive frames in the temporal detection frame sequence to represent the positions of corresponding points in the two consecutive frames. At the same time, the mask regions in the two consecutive temporal detection frames are fused by combining the positional changes (forward optical flow vector information sets and reverse optical flow vector information sets), so that the mask information of the first temporal detection frame has a certain effect on the second temporal detection frame, avoiding incomplete target recognition caused by masking only the content of a single frame. Combined with subsequent steps 1035-104, the fused mask is desensitized and the desensitized video image is output, reducing the amount of missing desensitization in some areas.

[0073] In addressing the aforementioned technical problems in the application scenario of desensitization for sports broadcast videos, the following technical issues often arise: If only forward optical flow is calculated during desensitization, and all vectors are directly used for coordinate mapping, a large number of false displacement vectors are generated when there are object occlusions, camera jumps, or missing textures in the video. These false vectors drive the mask coordinate mapping, leading to large-scale misalignment and random drift of the mask. Normal image areas are marked as sensitive areas to trigger desensitization, while areas occluded by moving targets may have missing masks, resulting in incomplete desensitization and poor desensitization performance. To address the following requirements for this application scenario: adaptability to situations such as object occlusion and camera jumps that frequently occur in sports broadcasts, and reduction of incomplete desensitization caused by missing masks, we have decided to adopt the following solution: In some optional implementations of certain embodiments, the aforementioned execution entity can perform mask deduction and fusion processing on the first temporal detection frame and the second temporal detection frame based on the desensitization mask information sequence, the forward optical flow vector information set, and the reverse optical flow vector information set to obtain a corrected fusion mask: The first step involves performing bidirectional optical flow consistency verification on the forward and reverse optical flow vector information sets to obtain a verification optical flow vector information set. In practice, for each pixel in the first temporal detection frame, the execution entity can perform the following steps: First, add the pixel's coordinates to the corresponding forward optical flow vector information; the resulting coordinates are used as the forward estimated coordinates. Next, subtract the forward estimated coordinates from the corresponding reverse optical flow vector information; the resulting coordinates are used as the reverse mapped coordinates. Then, the Euclidean distance between the pixel's coordinates and the reverse mapped coordinates is determined as the consistency error. Then, in response to determining that the consistency error is less than a preset error condition, the forward and reverse optical flow vector information corresponding to the pixel are determined as verification optical flow vector information. Finally, the obtained verification optical flow vector information is determined as the verification optical flow vector information set. The preset error condition can be that the consistency error is less than a preset value. Each set of verified optical flow vector information corresponds to a pixel in the first temporal detection frame and a pixel in the second temporal detection frame.

[0074] The second step is to determine the sensitive information mask corresponding to the first temporal detection frame as the target sensitive information mask. Each pixel in the target sensitive information mask corresponds to a verification optical flow vector in the aforementioned verification optical flow vector information set.

[0075] The third step involves performing a deduction process on the target sensitive information mask based on the verification optical flow vector information set, resulting in a target deduction mask. In practice, for each pixel in the aforementioned target sensitive information mask, the execution entity can perform the following steps: First, determine the verification optical flow vector information corresponding to the pixel as the target verification optical flow vector information. Next, determine the deduction coordinates by combining the coordinates of the pixel with the forward optical flow vector information in the target verification optical flow vector information, keeping the pixel mask value unchanged. Then, arrange each pixel in the aforementioned target sensitive information mask according to its corresponding deduction coordinates, keeping the pixel values ​​unchanged, and use the resulting image as the target deduction mask. It should be noted that if there are missing pixels in the rearranged target deduction mask image, the pixel mask value will be automatically filled with a preset value, 0 or 1.

[0076] The fourth step is to generate a filtering inference mask based on the target inference mask. In practice, firstly, the execution entity can determine the image size of the target sensitive information mask as the target size. Then, the target inference mask is cropped to the target size, and pixels outside the target size are directly deleted to obtain the cropped target inference mask as the filtering inference mask.

[0077] The fifth step involves generating motion intensity values ​​based on the forward optical flow vector information set and the first temporal detection frame. In practice, for each pixel in the first temporal detection frame, the execution entity can perform the following processing steps: First, the forward optical flow vector information corresponding to the pixel is determined as the target forward optical flow vector information. Next, the Euclidean distance formula is used to calculate the value of the target forward optical flow vector information as a single-point motion value. Then, the largest single-point motion value is determined as the motion intensity value.

[0078] The sixth step is to obtain the mask weight information corresponding to the motion intensity value. In practice, the execution entity can obtain the mask weight information corresponding to the motion intensity value from a preset storage device. This mask weight information can be pre-set weight information used for mask fusion. Different motion intensity values ​​can correspond to different mask weight information to adapt to different degrees of image changes.

[0079] Step 7: Determine the sensitive information mask corresponding to the second time-series detection frame as the sensitive information mask to be fused.

[0080] Step 8: Based on the mask weight information, perform weighted fusion processing on the filtered deduction mask and the sensitive information mask to be fused to obtain the fused mask. In practice, the above-mentioned execution entity can weight and superimpose the value of each pixel in the filtered deduction mask and the value of each pixel in the mask of the same position in the sensitive information to be fused according to the weight ratio in the mask weight information, and use the superimposed image as the fused mask.

[0081] The ninth step involves performing edge dilation correction on the fusion mask to obtain a corrected fusion mask. In practice, firstly, the execution entity can obtain the dilation radius corresponding to the motion intensity value from a preset storage device. Then, based on the dilation radius, morphological dilation operations are performed on the fusion mask to obtain the dilated fusion mask as the corrected fusion mask. The dilation radius can be the structuring element radius of the morphological dilation operation, in pixels; it controls how many pixels the mask expands outward.

[0082] The above technical solution, combined with subsequent steps 1035-104 and related content, serves as an inventive point of this disclosure, solving the technical problem of "poor desensitization effect." Factors leading to poor desensitization effects are often as follows: In the desensitization processing of sports broadcast videos, if only the forward optical flow is calculated, all vectors are directly used for coordinate mapping. When objects obstruct the video, there are camera jumps, or local texture loss, a large number of false displacement vectors are generated. These false vectors drive the mask coordinate mapping, causing large-area misalignment and random drift in the mask; normally visible areas are marked as sensitive areas to trigger desensitization, while areas obstructed by moving targets have missing masks, resulting in incomplete desensitization and poor desensitization effects. Solving these factors can improve the plate recognition effect. To achieve this effect, firstly, a bidirectional optical flow consistency verification process is performed on the forward and reverse optical flow vector information sets to obtain a verification optical flow vector information set. This allows verification of the forward and reverse optical flow vector information at the same point, discarding vector information with significant deviations. Next, the sensitive information mask corresponding to the first temporal detection frame is determined as the target sensitive information mask. Then, based on the verification optical flow vector information set, the target sensitive information mask is deduced to obtain the target deduced mask. Thus, the mask region in the first temporal detection frame can be forward deduced to obtain the mask information mapped from the first temporal detection frame to the second temporal detection frame. Then, based on the target deduced mask, a filtering deduced mask is generated. This removes points outside the mask region, reducing the impact of misaligned deduced points on subsequent fusion. Next, based on the forward and reverse optical flow vector information sets, motion intensity values ​​are generated. Thus, the abrupt change distance of the strongest abrupt change point in the image can be obtained as the motion intensity value. Then, the mask weight information corresponding to the motion intensity value is obtained. Thus, the weight information corresponding to mask fusion can be obtained. Finally, the sensitive information mask corresponding to the second temporal detection frame is determined as the sensitive information mask to be fused. Next, based on the mask weight information, the filtered derivation mask and the sensitive information mask to be fused are weighted and fused to obtain a fused mask. This yields a fused mask that combines mask information from the first and second temporal detection frames, reducing the dependence of the mask region on a single frame. Finally, edge dilation correction is applied to the fused mask to obtain a corrected fused mask. This expands the fused mask within a reasonable range to reduce mask loss at target edges. Furthermore, by first performing bidirectional verification on the forward and reverse optical flow vector information sets to identify and remove invalid displacement vectors caused by occlusion and sudden image changes, a verification optical flow vector information set is obtained. Then, using the verification optical flow vector information set, the sensitive information mask corresponding to the first temporal detection frame is deduced, and the mask information deduced from the action information is used as the target derivation mask.Then, invalid points (points outside the target area) in the target projection mask are filtered to reduce invalid masks caused by points with large offsets. Simultaneously, based on the intensity of image changes (motion intensity value), the filtered projection mask and the sensitive information mask to be fused are fused under different weights to obtain a fused mask. The edges of the fused mask are then expanded to reduce mask gaps. Combined with subsequent steps 1035-104, desensitization action information is generated based on the fused mask to desensitize the video image, reducing missed desensitization caused by mask gaps and improving the desensitization effect.

[0083] Step 1036: Based on the temporal detection frame sequence, update the video frame sequence to be detected and the detected video frame sequence to obtain the updated video frame sequence to be detected and the updated detected video frame sequence.

[0084] In some embodiments, the execution entity can update the video frame sequence to be detected and the detected video frame sequence based on the time-series detection frame sequence to obtain updated video frame sequences to be detected and updated detected video frame sequences. In practice, firstly, the execution entity can use each video frame in the video frame sequence to be detected whose timestamp falls within the time interval as an updated video frame, and arrange the updated video frames according to their chronological order in the video frame sequence to be detected to obtain an updated video frame sequence. Next, the updated video frames are deleted from the time-series detection frame sequence to update the detected video frame sequence. Then, the updated video frame sequence is added to the deleted portion of the detected video frame sequence to update the detected video frame sequence.

[0085] Step 1037: In response to determining that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence.

[0086] In some embodiments, the execution entity may, in response to determining that the updated video frame sequence to be detected is not empty, perform the above sensitive detection processing again on the updated video frame sequence to be detected based on the updated detected video frame sequence.

[0087] Step 104: In response to the determination that the updated video frame sequence to be detected is empty, based on the obtained desensitization action information, the detected video frame sequence is desensitized and reassembled to obtain the desensitized video source file and send the desensitized video source file to the preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0088] In some embodiments, the execution entity may, in response to determining that the updated video frame sequence to be detected is empty, perform desensitization and reassembly processing on the detected video frame sequence based on the obtained desensitization action information, obtain a desensitized video source file, and send the desensitized video source file to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0089] In some optional implementations of certain embodiments, the aforementioned execution entity may perform desensitization and reassembly processing on the aforementioned detected video frame sequence based on the obtained desensitization action information through the following steps to obtain the desensitized video source file: The first step is to determine the above-mentioned desensitization action information as a desensitization action information set.

[0090] The second step involves performing the following processing on the detected video frame sequence based on the desensitization action information set: The first sub-step involves extracting one desensitization action from the desensitization action information set as the target desensitization action information.

[0091] The second sub-step involves extracting target interval video frames from the detected video frame sequence based on the target desensitization action information, resulting in an interval video frame sequence. In practice, firstly, the executing entity can identify each detected video frame in the detected video frame sequence whose corresponding timestamp falls within the time interval corresponding to the action time interval in the target desensitization action information as an interval video frame. Then, the obtained interval video frames are arranged according to their chronological order within the detected video frame sequence to obtain the interval video frame sequence.

[0092] The third sub-step involves performing target region desensitization processing on the interval video frame sequence based on the target desensitization action information, resulting in a desensitized interval video frame sequence. In practice, firstly, the executing entity can determine the candidate bounding boxes corresponding to the sensitive target detection information in each sensitive target information set within the aforementioned target desensitization action information as target candidate bounding boxes. Then, according to the action type in the target desensitization action information, the regions corresponding to each interval video frame and each obtained target candidate bounding box in the aforementioned interval video frame sequence are processed according to the action type, resulting in a desensitized interval video frame sequence. This desensitized interval video frame sequence can be a sequence of video frames in which sensitive content is blurred or obscured in each frame. Each desensitized interval video frame in the aforementioned desensitized interval video frame sequence corresponds to one interval video frame in the aforementioned interval video frame sequence. As an example, in the above target desensitization action information, there is an action type of "very blurry," and {[Pedestrian 1: Box A - Frame 1, Box C - Frame 4, confidence level 0.88]}. Therefore, in all frames between frames 1 and 4, the positions of Box A and Box C will be blurred with the action corresponding to "very blurry," making the corresponding areas visually blurry. It should be noted that the time interval between frames 1 and 4 is short, the target movement distance is small, and the overlapping area of ​​two candidate images can generally cover the area where the target appears in frames 1 to 4. Optionally, in practice, the executing entity can also perform the following steps on each desensitization mask information in the updated desensitization mask information sequence in the above target desensitization action information: First, determine at least one interval video frame in the above interval video frame sequence whose corresponding timestamp falls within the mask time interval in the above desensitization mask information as at least one video frame to be desensitized. Then, for each of the at least one video frames to be desensitized, a preset desensitization action (blurring or mosaicking) is performed on the pixels in the video frame to be desensitized that correspond to the positions of the correction and fusion mask in the desensitization mask information, so as to complete the desensitization action on the interval video frame sequence and obtain the desensitized interval video frame sequence as the desensitized interval video frame sequence.

[0093] The fourth sub-step involves replacing the corresponding frames in the detected video frame sequence based on the anonymized interval video frame sequence, resulting in a replaced detected video frame sequence. In practice, the executing entity can replace the corresponding detected video frames in the detected video frame sequence with the anonymized interval video frame sequence to obtain the replaced detected video frame sequence.

[0094] The fifth sub-step involves deleting the target desensitization action information from the desensitization action information set to update the aforementioned desensitization action information set, thus obtaining the updated desensitization action information set.

[0095] The sixth sub-step, in response to determining that the updated desensitization action information set is not empty, involves performing the above processing again on the replaced detected video frame sequence based on the updated desensitization action information set.

[0096] The seventh sub-step, in response to the determination that the updated desensitization action information set is empty, recompresses and encodes the replaced detected video frame sequence to obtain the desensitized video source file. In practice, the executing entity can input the replaced detected video frame sequence into the FFmpeg tool to obtain the desensitized video source file. The desensitized video source file can be a video file obtained by compressing and encoding the replaced detected video frame sequence. The FFmpeg tool can be an open-source computer program licensed under the LGPL / GPL license.

[0097] The above-described embodiments of this disclosure have the following beneficial effects: the electronic screen sensitive image detection and desensitization methods of some embodiments of this disclosure reduce the abnormal playback of sensitive images. Specifically, the reason for the waste of computing resources and the occupation of data storage space is that: by performing frame-by-frame detection on the directly acquired video content to be played, and the detection process is performed on the sensitive content of a single frame, due to the large amount of time redundancy in the video stream (i.e., the content of multiple consecutive frames is highly similar), it will lead to repeated calculation of the same sensitive scene, which greatly wastes computing resources. If key frames are extracted for detection at fixed time intervals, although the computing load is reduced, it is very easy to miss short-duration sensitive images (such as "flickering" type violations) due to insufficient sampling frequency, or the detection mode may be maliciously circumvented due to the deviation in key frame selection, resulting in a serious risk of missed detection. Moreover, for single-frame detection, subsequent targeted desensitization operations are required for each image, generating a large amount of related desensitization action information, occupying a large amount of data storage space. Based on this, the electronic screen sensitive image detection and desensitization methods of some embodiments of this disclosure firstly, in response to receiving the detection signal sent by the preset electronic screen, acquire the video frame sequence to be detected. Therefore, the corresponding sequence of video frames to be detected for subsequent processing can be obtained. Next, the sequence of detected video frames is acquired. Then, based on the detected video frame sequence, the following sensitive detection processing is performed on the sequence of video frames to be detected: First, the first video frame to be detected in the sequence is identified as the target video frame. Next, based on the detected video frame sequence, frame duplication detection processing is performed on the target video frame to obtain frame duplication information. This allows us to determine whether the current video frame to be detected highly overlaps with previously processed video frames, reducing multiple detections of similar video frames. Then, in response to the determination that the frame duplication information does not meet the preset duplication condition, a temporal detection frame sequence is generated based on the frame duplication information, the target video frame, and the sequence of video frames to be detected. This allows us to generate adjacent video frames within a certain time window as a temporal detection frame sequence when the target video frame differs significantly from previously processed video frames. When there are significant changes in the image, more frames are extracted in a short time, reducing the possibility of missed detections. Finally, path-specific sensitive content detection processing is performed on the temporal detection frame sequence to obtain a sensitive detection information set. This allows us to obtain sensitive information from image semantics, target and text detection in the image. Next, based on the sensitive detection information set, desensitization action information is generated. This yields the desensitization actions to be performed on the target temporal detection frame sequence. Then, based on the temporal detection frame sequence, the video frame sequence to be detected and the already detected video frame sequence are updated to obtain the updated video frame sequence to be detected and the updated already detected video frame sequence.Next, in response to the determination that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence. Then, in response to the determination that the updated video frame sequence to be detected is empty, based on the obtained desensitization action information, the above detected video frame sequence is desensitized and reassembled to obtain a desensitized video source file, which is then sent to a preset electronic screen for output. This allows for the desensitization of the video images on the electronic screen. Furthermore, because the target video frame to be processed is first repeatedly detected along with all previously detected video frames, the waste of computational resources caused by detecting a large number of duplicate images is reduced. Next, by analyzing the differences between the target detection frame and previously detected video frames, the time interval and frequency of the extracted temporal detection frame sequence are determined, rather than using a single fixed value for extraction. This reduces the risk of missing short-duration sensitive frames (such as "flickering" violations) due to insufficient sampling frequency, or being maliciously circumvented by keyframe selection errors, thus minimizing the risk of missed detections. Simultaneously, a desensitization message is generated within a certain consecutive video frame interval (since the target position in most similar frames does not change significantly), reducing the generation of a large amount of related desensitization action information. This, in turn, reduces the waste of computing resources and the occupation of data storage space.

[0098] Further reference Figure 2 As an implementation of the methods shown in the figures, this disclosure provides some embodiments of a plate information recognition device, which are similar to... Figure 1 Corresponding to the method embodiments shown, the device can be specifically applied to various electronic devices.

[0099] like Figure 2As shown, an electronic screen sensitive image detection and desensitization device 200 in some embodiments includes: a first acquisition unit 201, a second acquisition unit 202, a detection processing unit 203, and a sending unit 204. The first acquisition unit 201 is configured to acquire a sequence of video frames to be detected in response to receiving a detection signal sent by a preset electronic screen; the second acquisition unit 202 is configured to acquire a sequence of detected video frames; the detection processing unit 203 is configured to perform the following sensitive detection processing on the sequence of video frames to be detected based on the detected video frame sequence: determining the first video frame to be detected in the sequence of video frames to be detected as the target video frame; performing frame repetition detection processing on the target video frame based on the detected video frame sequence to obtain frame repetition information; in response to determining that the frame repetition information does not meet a preset repetition condition, generating a time-series detection frame sequence based on the frame repetition information, the target video frame, and the sequence of video frames to be detected; and performing path-specific sensitive content detection processing on the time-series detection frame sequence to obtain sensitive detection information. The system generates a set of sensitive detection information; based on the sensitive detection information set, it generates desensitization action information; based on the time-series detection frame sequence, it updates the video frame sequence to be detected and the detected video frame sequence to obtain an updated video frame sequence to be detected and an updated detected video frame sequence; in response to determining that the updated video frame sequence to be detected is not empty, it performs the above sensitive detection processing again on the updated video frame sequence to be detected based on the updated detected video frame sequence; the sending unit 204 is configured to, in response to determining that the updated video frame sequence to be detected is empty, perform desensitization and reassembly processing on the detected video frame sequence based on the obtained desensitization action information, obtain a desensitized video source file, and send the desensitized video source file to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0100] It is understandable that the units described in the electronic screen sensitive image detection and desensitization device 200 are related to the reference. Figure 1 The steps in the method described above correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 200 and the units contained therein, and will not be repeated here.

[0101] The following is for reference. Figure 3 It shows a schematic diagram of the structure of an electronic device 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0102] like Figure 3As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory 302 or a program loaded from a storage device 308 into a random access memory 303. The random access memory 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, the read-only memory 302, and the random access memory 303 are interconnected via a bus 304. An input / output interface 305 is also connected to the bus 304.

[0103] Typically, the following devices can be connected to the input / output interface 305: input devices 306 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 307 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 308 including, for example, magnetic tape, hard disk, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0104] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a storage device 308, or installed from a read-only memory 302. When the computer program is executed by the processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0105] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical fiber, a portable compact disk read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0106] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0107] The computer-readable medium may be included in an electronic device or may exist independently without being assembled into the electronic device. The computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: in response to receiving a detection signal sent by a preset electronic screen, acquire a sequence of video frames to be detected; acquire a sequence of detected video frames; and, based on the detected video frame sequence, perform the following sensitive detection processing on the sequence of video frames to be detected: determine the first video frame to be detected in the sequence of video frames to be detected as the target video frame; perform frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information; in response to determining that the frame duplication information does not meet a preset duplication condition, generate a time-series detection frame sequence based on the frame duplication information, the target video frame, and the sequence of video frames to be detected; and perform path-specific sensitive content detection processing on the time-series detection frame sequence. The process involves obtaining a sensitive detection information set; generating desensitization action information based on the sensitive detection information set; updating the video frame sequence to be detected and the detected video frame sequence based on the time-series detection frame sequence to obtain an updated video frame sequence to be detected and an updated detected video frame sequence; in response to determining that the updated video frame sequence to be detected is not empty, performing the above sensitive detection process again on the updated detected video frame sequence based on the updated detected video frame sequence; and in response to determining that the updated video frame sequence to be detected is empty, performing desensitization and reloading processing on the detected video frame sequence based on the obtained desensitization action information to obtain a desensitized video source file and sending the desensitized video source file to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

[0108] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0109] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0110] The units described in some embodiments of this disclosure can be implemented in software or in hardware. The described units can also be located in a processor.

[0111] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0112] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of technical features, but should also cover other technical solutions formed by arbitrary combinations of technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for detecting and desensitizing sensitive images on an electronic screen, characterized in that, include: In response to receiving a detection signal sent by a preset electronic screen, the sequence of video frames to be detected is acquired; Obtain the sequence of detected video frames; Based on the detected video frame sequence, the following sensitive detection processing is performed on the video frame sequence to be detected: The first video frame to be detected in the sequence of video frames to be detected is identified as the target video frame. Based on the detected video frame sequence, frame duplication detection processing is performed on the target video frame to obtain frame duplication information; In response to the determination that the frame repetition information does not meet the preset repetition condition, a temporal detection frame sequence is generated based on the frame repetition information, the target video frame, and the video frame sequence to be detected. The temporal detection frame sequence is processed by path-specific sensitive content detection to obtain a sensitive detection information set; Based on the sensitive detection information set, generate desensitization action information; Based on the temporal detection frame sequence, the video frame sequence to be detected and the detected video frame sequence are updated to obtain the updated video frame sequence to be detected and the updated detected video frame sequence. In response to the determination that the updated video frame sequence to be detected is not empty, the above sensitive detection processing is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence. In response to the determination that the updated video frame sequence to be detected is empty, based on the obtained desensitization action information, the detected video frame sequence is desensitized and reassembled to obtain a desensitized video source file and the desensitized video source file is sent to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

2. The method according to claim 1, characterized in that, Each video frame in the sequence of video frames to be detected has a corresponding timestamp, and a temporal detection frame sequence is generated based on frame repetition information, the target video frame, and the sequence of video frames to be detected, including: Obtain the timing extraction information corresponding to the frame repetition information, wherein the timing extraction information includes the extracted frame value and the extraction time window information; Based on the extracted time window information, the time window video frame extraction process is performed on the video frame sequence to be detected to obtain the time window video frame sequence. Based on the extracted frame values, dynamic frame extraction is performed on the video frame sequence within the time window to obtain the time-series detection frame sequence.

3. The method according to claim 1, characterized in that, The process of performing path-specific sensitive content detection on the temporal detection frame sequence yields a sensitive detection information set, including: The following processing is performed on each timing detection frame in the timing detection frame sequence: Preprocess the timing detection frames to obtain preprocessed video frames; Image semantic category detection is performed on the preprocessed video frames to obtain image semantic category information; The pre-processed video frames are input into a pre-trained sensitive target detection model to obtain a sensitive target detection information set; Sensitive text recognition processing is performed on preprocessed video frames to obtain sensitive text information; Image semantic category information, sensitive target detection information set, and sensitive text information are identified as sensitive detection information; The obtained sensitive detection information is defined as a sensitive detection information set.

4. The method according to claim 3, characterized in that, The pre-trained sensitive target detection model includes a backbone network, a neck network, and a detection head. The process of inputting pre-processed video frames into the pre-trained sensitive target detection model to obtain a sensitive target detection information set includes: The preprocessed video frames are input into the backbone network to obtain the first feature fusion map; The first feature fusion map is input into the neck network to perform feature enhancement processing on the first feature fusion map, resulting in the first feature enhancement map; The first feature enhancement map is input into the detection head to perform multi-scale feature detection processing on the first feature enhancement map to obtain the feature detection map set; After applying nonmaximum suppression to the feature detection map, a sensitive target detection information set is obtained.

5. The method according to claim 3, characterized in that, The generation of desensitization action information based on the sensitive detection information set includes: The sensitive detection information set is normalized and mapped to obtain a normalized mapped information set. Based on the normalized mapping information set, multi-frame mapping information is generated; In response to the multi-frame mapping information satisfying the preset multi-frame mapping conditions, the following steps are performed on each sensitive detection information in the sensitive detection information set: The sensitive target detection information set in the sensitive detection information is processed to obtain the sensitive target information set; The sensitive text information in the sensitive detection information is processed to locate the text, resulting in a text location information set. Based on image semantic category information, sensitive target information set, and text location information set, generate single-frame desensitization action information; The desensitization action information of each single frame is processed into desensitized image intervals to obtain desensitization action information.

6. The method according to claim 1, characterized in that, Based on the obtained desensitization action information, the detected video frame sequence is desensitized and reassembled to obtain the desensitized video source file, including: Each desensitization action information is defined as a desensitization action information set; Based on the desensitized action information set, the following processing is performed on the detected video frame sequence: Extract one desensitization action from the desensitization action information set as the target desensitization action information; Based on the target desensitization action information, the detected video frame sequence is processed by target interval video frame extraction to obtain the interval video frame sequence. Based on the target desensitization action information, the target region desensitization processing is performed on the interval video frame sequence to obtain the desensitized interval video frame sequence; Based on the desensitized interval video frame sequence, the corresponding frames of the detected video frame sequence are replaced to obtain the replaced detected video frame sequence. The target desensitization action information is deleted from the desensitization action information set in order to update the desensitization action information set and obtain the updated desensitization action information set. In response to the determination that the updated desensitization action information set is not empty, the above processing is performed again on the replaced detected video frame sequence based on the updated desensitization action information set; In response to the determination that the updated desensitization action information set is empty, the replaced detected video frame sequence is recompressed and encoded to obtain the desensitized video source file.

7. A device for detecting and desensitizing sensitive images on an electronic screen, characterized in that, include: The first acquisition unit is configured to acquire a sequence of video frames to be detected in response to receiving a detection signal sent by a preset electronic screen; The second acquisition unit is configured to acquire the detected video frame sequence; The detection processing unit is configured to perform the following sensitive detection processing on the video frame sequence to be detected based on the detected video frame sequence: determining the first video frame to be detected in the video frame sequence to be detected as the target video frame; performing frame duplication detection processing on the target video frame based on the detected video frame sequence to obtain frame duplication information; in response to determining that the frame duplication information does not meet the preset duplication condition, generating a temporal detection frame sequence based on the frame duplication information, the target video frame, and the video frame sequence to be detected; and performing path-specific sensitive content detection processing on the temporal detection frame sequence to obtain a sensitive detection information set. Based on the sensitive detection information set, generate desensitization action information; Based on the temporal detection frame sequence, the video frame sequence to be detected and the detected video frame sequence are updated to obtain the updated video frame sequence to be detected and the updated detected video frame sequence. In response to the determination that the updated video frame sequence to be detected is not empty, the above sensitive detection process is performed again on the updated video frame sequence to be detected based on the updated detected video frame sequence. The sending unit is configured to, in response to determining that the updated video frame sequence to be detected is empty, perform desensitization and reassembly processing on the detected video frame sequence based on the obtained desensitization action information, obtain a desensitized video source file, and send the desensitized video source file to a preset electronic screen for the preset electronic screen to output the desensitized video source file.

8. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 6.

Images