Artificial intelligence-based neonatal image record detection method, system, device and medium

Abstract

The application relates to a new-born image record detection method, system, device and medium based on artificial intelligence. The method comprises the following steps: completing identity binding through wristband near field communication, generating a dynamic blood relationship label and embedding the video stream, combining multi-modal sensing data, identifying key events through a time sequence convolution network and an obstetric knowledge graph, and realizing intelligent control of image acquisition; a Mask R-CNN model is used to identify sensitive areas, and a segmented Logistic chaotic permutation based on UID and a timestamp is used to realize irreversible desensitization, and finally, the event is cut and archived according to the event and verified and uploaded. The method intelligently identifies key events of a new-born through multi-modal sensing data combined with an obstetric knowledge graph, realizes irreversible covering of sensitive areas through segmented Logistic chaotic pixel permutation, and realizes structured archiving based on events, thereby effectively improving the identity security, event identification accuracy, privacy protection strength and data management standardization of new-born image records, and guaranteeing the traceability and clinical reuse value of data.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, device and medium for newborn image recording and detection based on artificial intelligence. Background Technology

[0002] Neonatal imaging is a crucial component of modern obstetric and maternal-infant services and standardized medical data management. Accurately recording imaging data of key physiological events such as a newborn's first cry, touch, and eye opening not only provides direct, original evidence for retrospective analysis of neonatal growth and development but also accumulates standardized sample data for obstetric clinical research. Furthermore, it is a core element in enhancing the maternal and infant service experience. Therefore, the need for automated and intelligent acquisition and recording of neonatal images in hospital obstetrics wards is increasingly urgent.

[0003] However, current neonatal imaging recording technologies and applications still face numerous technical challenges that urgently need to be addressed, making it difficult to meet the practical application requirements of obstetric clinical practice. On the one hand, existing neonatal image acquisition largely relies on medical staff manually activating the acquisition equipment. However, obstetric medical staff have heavy daily workloads, making it easy to miss recordings of critical neonatal events, resulting in low critical event capture rates. Furthermore, while some technologies attempt to use simple sensor threshold triggering methods, they cannot accurately distinguish between valid neonatal physiological events and ward environment interference, easily triggering invalid acquisitions and generating a large amount of redundant video data. This significantly increases data storage costs and the workload of subsequent editing and processing, leading to low data resource utilization. On the other hand, there is a lack of a secure and reliable source binding mechanism between image data and neonatal identity. Existing binding technologies mostly rely on manual association of neonatal identity information after image acquisition, which easily leads to confusion and mismatch between image files and neonatal identities. This does not meet the high standards of standardized management for medical data traceability and accountability, and also poses serious risks to the establishment and subsequent use of neonatal image files.

[0004] Furthermore, during the acquisition of newborn images, the images inevitably contain sensitive and private information such as the faces of medical staff, department signs, and other infants in the ward background. Existing privacy protection methods mostly rely on manual screening and processing of sensitive areas, which is not only extremely inefficient but also has a high false positive rate. Some privacy filtering schemes that use general deep learning models lack specific adaptation to the obstetrics-specific scenario, cannot accurately identify various sensitive areas in the obstetrics scenario, and have poor privacy protection effects. They are also prone to over-covering the effective newborn image area or omitting the coverage of sensitive areas, posing a serious risk of privacy leakage. Summary of the Invention

[0005] Therefore, it is necessary to provide artificial intelligence-based methods, systems, devices, and media for newborn image recording and detection to address the aforementioned technical issues. This aims to enhance the security and traceability of binding image data with newborn identity, improve the accuracy and processing efficiency of image privacy protection in obstetric settings, reduce the generation of redundant image data, and achieve structured management and secure archiving of newborn image data.

[0006] Firstly, this application provides an artificial intelligence-based method for detecting neonatal image recordings, including:

[0007] When the near-field communication signal of the wristband is detected, the UID and basic information of the newborn in the wristband are read. The initial video stream is acquired in real time through the image acquisition device. The current system timestamp is obtained as the binding timestamp. A dynamic bloodline tag is generated based on the UID and the binding timestamp. The dynamic bloodline tag is embedded into the initial image of each frame of the initial video stream through frequency domain transformation to obtain the bloodline binding video stream. The UID, binding timestamp, dynamic bloodline tag and basic information of the newborn are packaged to generate a global context object.

[0008] The system acquires audio and accelerometer data streams in real time, extracts features from these streams to obtain multimodal feature tensors, inputs these tensors into a pre-defined temporal convolutional network, and outputs event probability vectors. Based on these event probability vectors, it queries an obstetrics knowledge graph for rule-based reasoning. When the reasoning result confirms an event, event information is generated. A data acquisition trigger command is generated based on the event information, and the event information is stored in a pre-defined event list. This pre-defined event list is then appended to a global context object to obtain an optimized global context object. The data acquisition trigger command controls the start and stop of the initial video stream acquisition by the image acquisition device. Event information includes event type, start timestamp, and estimated end time.

[0009] For each frame of the bloodline-bound video stream, the timestamp corresponding to the frame image is obtained as the current frame timestamp. The frame image is input into a preset Mask R-CNN instance segmentation network to identify sensitive regions that are not newborn subjects. The desensitization strategy is dynamically adjusted according to the event type in the optimized global context object. For sensitive regions, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. The sensitive regions are irreversibly covered by a chaotic pixel replacement algorithm to generate desensitized video frames. The sensitive region categories and processing parameters during the desensitization process are recorded to generate desensitization operation metadata. The desensitization operation metadata is appended to the optimized global context object to obtain the final global context object. Combined with each desensitized video frame, the desensitized video stream is obtained.

[0010] The system reads a list of preset events from the final global context object. Based on the start timestamp and estimated end time of each event in the preset event list, it divides the anonymized video stream into multiple video segments along the timeline. It generates a metadata file for each video segment and extracts embedded dynamic lineage tags from the video segments. The embedded dynamic lineage tags are compared and verified with the dynamic lineage tags in the final global context object to obtain the tag verification results. If the tag verification results are consistent, the system packages each video segment and its corresponding metadata file into a structured archive and uploads the structured archive to the hospital information system.

[0011] In one embodiment, a dynamic lineage tag is generated based on the UID and the binding timestamp, and the dynamic lineage tag is embedded into the initial image of each frame of the initial video stream through frequency domain transformation to obtain a lineage-bound video stream, including:

[0012] A hash algorithm is used to calculate the bound timestamp to obtain a 256-bit dynamic factor; the UID and the dynamic factor are then XORed bitwise to generate a 128-bit dynamic lineage tag.

[0013] The luminance component of each frame of the initial video stream is divided into 8×8 blocks, and a discrete cosine transform is performed on each block to obtain the frequency domain coefficient matrix of each block.

[0014] The dynamic lineage tag is converted into a binary bit stream. The intermediate frequency coefficients of each block are selected as the embedding carrier according to the Zigzag scanning order. By adjusting the numerical characteristics of the intermediate frequency coefficients, the corresponding bits of the binary bit stream are embedded into the intermediate frequency coefficients of each block one by one, and the embedded block frequency domain coefficient matrix is ​​obtained.

[0015] Perform inverse discrete cosine transform on each embedded block frequency domain coefficient matrix to obtain the pixel matrix of each block;

[0016] The pixel matrices of each block are spliced ​​and integrated according to the original image block positions during the block segmentation process to obtain the image frame with embedded tags;

[0017] The image frames after each embedded tag are combined sequentially in chronological order to obtain the bloodline-bound video stream.

[0018] In one embodiment, the preset temporal convolutional network includes a feature extraction module, a residual module, a global average pooling layer, and a fully connected layer, wherein the residual module is constructed based on an attention gating mechanism; the accelerometer data stream is triaxial acceleration data.

[0019] Feature extraction is performed on the audio stream and the accelerometer sensor data stream to obtain a multimodal feature tensor; the multimodal feature tensor is input into a pre-defined temporal convolutional network to output an event probability vector, including:

[0020] The audio stream is segmented into frames. Based on the preset frame length and frame shift, a Hamming window is applied to each frame of audio. The first 13-dimensional Mel frequency cepstral coefficients are calculated. The single-frame audio feature vector is obtained by combining short-time energy and zero-crossing rate. The single-frame audio feature vectors of consecutive preset frames are stacked to obtain the audio feature tensor.

[0021] Based on the preset window length and step size, a sliding window is used to process the acceleration sensor data stream. The mean, variance, peak-to-peak value, energy entropy and spectral peak features are extracted from the triaxial acceleration data in each sliding window to generate the acceleration feature vector corresponding to each sliding window. The acceleration feature vectors of multiple consecutive sliding windows are stacked to obtain the acceleration feature tensor.

[0022] The audio feature tensor and the acceleration feature tensor are time-aligned to obtain the multimodal feature tensor;

[0023] The multimodal feature tensor is input into a preset temporal convolutional network. The feature extraction module extracts temporal features from the multimodal feature tensor to obtain initial temporal features. The residual module calculates the gate coefficients. The initial temporal features are then weighted and enhanced based on the gate coefficients to obtain weighted temporal features.

[0024] The weighted temporal features are reduced in dimensionality by using a global average pooling layer of a pre-defined temporal convolutional network. The reduced features are then input into a fully connected layer for classification operations, and the output is an event probability vector containing the probabilities of crying, touching, opening eyes, and no events.

[0025] In one embodiment, rule-based reasoning is performed by querying the obstetric knowledge graph based on the event probability vector. When the reasoning result confirms that an event has occurred, event information is generated, including:

[0026] Obtain a preset first threshold and a second threshold. When the probability of any event in the event probability vector exceeds the first threshold, the event whose probability exceeds the first threshold is identified as a candidate event.

[0027] Extract the timestamp, audio energy parameter, and peak acceleration parameter corresponding to the candidate event as context attributes, and construct a query graph based on the context attributes;

[0028] Input the query graph into the preset obstetrics knowledge graph for matching query to obtain entity association matching results; the obstetrics knowledge graph includes entities, attributes and preset medical common sense defined in the form of rules. Entities include newborns, various target events and image acquisition equipment. Attributes include event duration and event association relationship. Medical common sense includes event duration rules and event association verification rules.

[0029] By using the rule-based reasoning engine of the obstetrics knowledge graph and combining entity association matching results, we can verify whether candidate events conform to preset medical common sense and output the event confidence adjustment factor.

[0030] The event confidence adjustment factor is multiplied by the event probability corresponding to the candidate event in the event probability vector to obtain the final confidence and generate the corresponding inference result.

[0031] If the final confidence level exceeds the second threshold, the inference result is that the event has occurred. The event type is determined according to the type of the candidate event. The center time of the feature extraction window corresponding to the candidate event is used as the start timestamp. The estimated end time is determined based on the event duration rules in the obstetrics knowledge graph. The feature extraction window is the audio frame window corresponding to the Hamming window or the acceleration feature extraction window corresponding to the sliding window.

[0032] Integrate event type, start timestamp, and estimated end time to generate corresponding event information.

[0033] In one embodiment, the preset Mask R-CNN instance segmentation network includes a region proposal network and a classification branch. The region proposal network includes 16×16 anchor boxes and 32×32 anchor boxes. The frame image is input into the preset Mask R-CNN instance segmentation network to identify sensitive regions that are not newborn subjects. The desensitization strategy is dynamically adjusted based on the event type in the optimized global context object. For sensitive regions, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. A chaotic pixel replacement algorithm is used to irreversibly mask the sensitive regions, generating desensitized video frames. The sensitive region categories and processing parameters during the desensitization process are recorded, generating desensitization operation metadata, including:

[0034] The frame image is input into the preset Mask R-CNN instance segmentation network. Multiple candidate region boxes are generated through the 16×16 and 32×32 anchor boxes of the region proposal network. The region of interest pooling operation is performed on the candidate region boxes to obtain the feature map corresponding to each candidate region box.

[0035] The feature maps are input into the classification branch of the preset Mask R-CNN instance segmentation network. The classification branch determines the category of the target corresponding to each feature map and predicts the segmentation mask and confidence level of each target. The bounding box, category label, segmentation mask and confidence level of each target are output. Targets with confidence levels exceeding the preset confidence threshold are identified as detection targets. Targets with category labels of medical staff faces, department logos or background infants are selected as sensitive regions from the detection targets.

[0036] The event type is extracted from the optimized global context object, and the desensitization strategy is dynamically adjusted according to the event type to obtain the target desensitization strategy; the UID in the optimized global context object is concatenated with the current frame timestamp to obtain the concatenation result, and a hash operation is performed on the concatenation result to obtain the initial seed of the chaotic mapping;

[0037] For each sensitive region, the initial seed of the chaotic mapping is iteratively calculated based on the Logistic chaotic mapping formula to generate a chaotic sequence equal to the total number of pixels in the sensitive region, and a permutation index array equal to the total number of pixels in the sensitive region is generated based on the chaotic sequence.

[0038] The pixels in the sensitive area are expanded into a one-dimensional pixel array in row order. The pixel positions of the one-dimensional pixel array are rearranged according to the permutation index array to obtain the rearranged one-dimensional pixel array.

[0039] The rearranged one-dimensional pixel array is filled back into the sensitive area according to the original area coordinates. The filled sensitive area is then merged with the unprocessed area in the frame image to generate a desensitized video frame.

[0040] Record the category of the sensitive region, the bounding box coordinates, the initial seed of the chaotic mapping, and the generation parameters of the permutation index array to obtain the metadata of the desensitization operation.

[0041] In one embodiment, the mathematical expression of the Logistic chaotic mapping formula is:

[0042]

[0043]

[0044]

[0045] in, Use it as the initial seed for chaotic mapping; For the UID Bit binary value, ; The timestamp of the current frame Bit binary value, ; This is a bitwise XOR operation; For the first The values ​​of the chaotic sequence in the next iteration; and All are piecewise chaos control parameters; and This refers to the category weight adjustment coefficient for sensitive areas; Weights for sensitive region categories; For the first pixel displacement index; This represents the total number of pixels in the sensitive area.

[0046] Secondly, this application also provides an artificial intelligence-based neonatal image recording and detection system, including:

[0047] The identity binding and video acquisition module is used to read the UID and newborn's basic information from the wristband when the near-field communication signal of the wristband is detected. It then acquires the initial video stream in real time through the image acquisition device, obtains the current system timestamp as the binding timestamp, generates a dynamic bloodline tag based on the UID and binding timestamp, and embeds the dynamic bloodline tag into the initial image of each frame of the initial video stream through frequency domain transformation to obtain the bloodline binding video stream. Finally, it packages the UID, binding timestamp, dynamic bloodline tag and newborn's basic information into a global context object.

[0048] The intelligent event triggering module is used to acquire audio streams and accelerometer data streams in real time, extract features from these streams to obtain multimodal feature tensors, input these tensors into a preset temporal convolutional network, and output event probability vectors. Based on these event probability vectors, the module queries the obstetrics knowledge graph for rule-based reasoning. When the reasoning result confirms an event has occurred, event information is generated. A collection trigger command is generated based on the event information, and the event information is stored in a preset event list. This preset event list is then appended to a global context object to obtain an optimized global context object. The collection trigger command controls the start and stop of the initial video stream acquisition by the image acquisition device. The event information includes the event type, start timestamp, and estimated end time.

[0049] The privacy protection and desensitization module is used to obtain the timestamp corresponding to each frame of the bloodline-bound video stream as the current frame timestamp. The frame image is then input into a preset Mask R-CNN instance segmentation network to identify sensitive areas that are not newborn subjects. The desensitization strategy is dynamically adjusted according to the event type in the optimized global context object. For sensitive areas, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. The sensitive areas are irreversibly masked using a chaotic pixel replacement algorithm to generate desensitized video frames. The module also records the sensitive area categories and processing parameters during the desensitization process, generates desensitization operation metadata, and appends the desensitization operation metadata to the optimized global context object to obtain the final global context object. Combined with each desensitized video frame, the desensitized video stream is obtained.

[0050] The document structuring and archiving module is used to read a preset event list from the final global context object, and divide the desensitized video stream into multiple video segments along the timeline according to the start timestamp and estimated end time of each event in the preset event list; generate a metadata file for each video segment, and extract embedded dynamic lineage tags from the video segments; compare and verify the embedded dynamic lineage tags with the dynamic lineage tags in the final global context object to obtain the tag verification result; if the tag verification result is consistent, package each video segment and the corresponding metadata file into a structured archive package, and upload the structured archive package to the hospital information system.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0053] The aforementioned AI-based neonatal image recording and detection method, system, equipment, and media firstly, utilize near-field communication to instantly read the UID and generate a dynamic lineage tag embedded in the video frame in the frequency domain, solving the problems of identity confusion and difficulty in tracing caused by traditional post-event manual binding. Secondly, by fusing audio and accelerometer sensor data through temporal convolutional network inference, it achieves accurate identification and automated acquisition control of key physiological events in newborns, eliminating reliance on manual initiation and reducing false alarms due to environmental interference. Furthermore, the rule-based reasoning mechanism driven by obstetric knowledge graphs compensates for the limitations of threshold judgment by single sensors, improving the accuracy and reliability of event recognition. Finally, by dynamically adjusting the desensitization strategy based on event type and combining it with chaotic pixel replacement, it solves the problems of insufficient desensitization accuracy and privacy leakage risks of general models, enhancing the privacy protection effectiveness in obstetric-specific scenarios. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart of an artificial intelligence-based neonatal image recording and detection method is provided as an exemplary embodiment of the present invention;

[0056] Figure 2A flowchart illustrating a method for obtaining an event probability vector, provided as an exemplary embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of a newborn image recording and detection system based on artificial intelligence, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based method for newborn image recording and detection is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S101: When the near-field communication signal of the wristband is detected, the UID and basic information of the newborn in the wristband are read. The initial video stream is acquired in real time through the image acquisition device. The current system timestamp is obtained as the binding timestamp. A dynamic bloodline tag is generated based on the UID and the binding timestamp. The dynamic bloodline tag is embedded into the initial image of each frame of the initial video stream through frequency domain transformation to obtain the bloodline binding video stream. The UID, binding timestamp, dynamic bloodline tag and basic information of the newborn are packaged to generate a global context object.

[0061] Specifically, when the near-field communication signal of the wristband is detected, the source binding operation of newborn identification and image identity can be performed. The wristband is a disposable medical encrypted wristband specifically for obstetrics and gynecology, with an anti-tamper structure design, a built-in 13.56MHz RFID chip, and an encrypted QR code bound to the chip's UID printed on the surface. The wristband itself is equipped with an anti-tamper ring structure, which will become invalid once forcibly disassembled, thus effectively avoiding the risk of the wristband being reused and identity information being misused.

[0062] In a demonstrative manner, the unique UID stored in the wristband chip can be read via 13.56MHz high-frequency near-field communication, and a secondary verification of identity information can be performed by scanning the encrypted QR code on the surface. The read UID, together with the newborn's basic information, constitutes a unique identifier for the newborn, laying the foundation for a strong association between subsequent images and identity. Simultaneously, image acquisition devices deployed around the crib, such as multimodal AI acquisition devices, can capture the initial video stream in real time. The current system timestamp can be used as the binding timestamp. Utilizing the uniqueness of the UID and the dynamism of the timestamp, a dynamic lineage tag is generated to avoid tag duplication and enhance the security of identity binding. Embedding this dynamic lineage tag into each initial frame of the initial video stream through frequency domain transformation achieves a physical binding between the image frame and identity information, rather than a simple file-level association. This eliminates the problems of image-identity separation and file confusion from the data source, ultimately resulting in a lineage-bound video stream with the characteristic of unalterable identity information. By packaging the UID, binding timestamp, dynamic lineage tag, and newborn's basic information into a global context object, unified management of newborn identity-related data can be achieved. Furthermore, this object serves as the core data carrier of the entire method, providing unified and efficient information support for identity association operations in subsequent steps.

[0063] S102: Real-time acquisition of audio stream and accelerometer sensor data stream; feature extraction of audio stream and accelerometer sensor data stream to obtain multimodal feature tensors; inputting multimodal feature tensors into a preset temporal convolutional network to output event probability vectors; querying the obstetrics knowledge graph based on event probability vectors for rule-based reasoning; when the reasoning result confirms the occurrence of an event, generating event information; generating acquisition trigger commands based on event information; storing event information in a preset event list; and appending the preset event list to a global context object to obtain an optimized global context object. The acquisition trigger commands are used to control the start and stop of the image acquisition device's acquisition of the initial video stream; event information includes event type, start timestamp, and estimated end time.

[0064] Specifically, after binding the identity and image source, audio streams and accelerometer data streams synchronously acquired by the multimodal AI acquisition device can be obtained in real time. The audio stream is used to capture key acoustic events such as newborn crying, while the accelerometer data stream is used to capture key motor events such as newborn limb movements and touch. Feature extraction is performed on these two types of raw unstructured sensor data streams, which can be converted into structured multimodal feature tensors that can be recognized by artificial intelligence models, realizing feature quantification of sensor data. Subsequently, the multimodal feature tensors can be input into a preset temporal convolutional network. This network can be adapted and optimized for neonatal obstetric scenarios, accurately mining temporal correlation features in multimodal data and outputting event probability vectors containing various key neonatal events. The obstetric knowledge graph integrates obstetric clinical medical knowledge with neonatal critical event determination rules. Rule-based reasoning based on this probability vector query of the obstetric knowledge graph compensates for the lack of medical basis in pure model reasoning, improving the scientific rigor and accuracy of event determination. When the reasoning result confirms an event, event information is generated, and a collection trigger command is generated based on this information. This command can control the start and stop of the initial video stream acquisition by the image acquisition device, achieving intelligent triggering of image acquisition, replacing traditional manual triggering methods, effectively avoiding missed acquisition of neonatal critical events. Furthermore, high-definition acquisition can be initiated only when a confirmed event occurs, significantly reducing redundant data generated by continuous recording and lowering data storage and subsequent processing costs. Storing event information in a preset event list and appending this list to a global context object to obtain an optimized global context object enables unified management of neonatal critical event information and identity information, providing precise event data support for subsequent video segmentation and dynamic adjustment of desensitization strategies.

[0065] S103: For each frame of the bloodline-bound video stream, obtain the timestamp corresponding to the frame image as the current frame timestamp, input the frame image into a preset Mask R-CNN instance segmentation network to identify sensitive areas that are not newborn subjects; dynamically adjust the desensitization strategy according to the event type in the optimized global context object; for sensitive areas, generate a chaotic sequence based on the UID in the optimized global context object and the current frame timestamp, irreversibly cover the sensitive areas through a chaotic pixel replacement algorithm, generate desensitized video frames, and record the sensitive area category and processing parameters during the desensitization process to generate desensitization operation metadata; append the desensitization operation metadata to the optimized global context object to obtain the final global context object; combine each desensitized video frame to obtain the desensitized video stream.

[0066] Specifically, for each frame of a bloodline-bound video stream, its corresponding timestamp can be obtained first as the current frame timestamp, achieving a precise correlation between the frame image and the time dimension, providing a temporal basis for the dynamic adjustment of the desensitization strategy. The frame image is then input into a preset Mask R-CNN instance segmentation network. This network can accurately identify sensitive areas in obstetric scenes that are not primarily newborn subjects, such as the faces of medical staff, department logos, and background infants. This avoids including newborn subjects in the desensitization scope from a logical perspective, ensuring the integrity of the newborn image. Furthermore, the desensitization strategy can be dynamically adjusted based on the event type in the optimized global context object, adapting privacy desensitization operations to different key newborn event scenarios, improving the targeting and rationality of the desensitization process. Subsequently, based on the adjusted desensitization strategy, for the identified sensitive areas, a chaotic sequence can be generated based on the UID and the current frame timestamp in the optimized global context object. Utilizing the uniqueness of the UID and the dynamic nature of the current frame timestamp, the generated chaotic sequence possesses unpredictable and unique characteristics. Based on this chaotic sequence, an irreversible masking algorithm is used to apply the chaotic pixel replacement algorithm to the sensitive areas. Compared to traditional image blurring methods, the aforementioned irreversible pixel replacement operation effectively enhances privacy protection and prevents sensitive information from being restored. The generated desensitized video frames effectively filter sensitive privacy information of non-newborn subjects while fully preserving the image information of the newborn. Furthermore, it can record the categories of sensitive areas and processing parameters during the desensitization process, generating desensitization operation metadata. This metadata is appended to the optimized global context object to obtain the final global context object, achieving unified management of desensitized information, identity, and event information. The desensitized video stream obtained by combining all desensitized video frames possesses the dual characteristics of identity binding and privacy desensitization, complying with medical data privacy protection standards.

[0067] S104: Read the preset event list from the final global context object, and divide the desensitized video stream into multiple video segments along the timeline according to the start timestamp and estimated end time of each event in the preset event list; generate a metadata file for each video segment, and extract the embedded dynamic lineage tags from the video segments; compare and verify the embedded dynamic lineage tags with the dynamic lineage tags in the final global context object to obtain the tag verification result; if the tag verification result is consistent, package each video segment and the corresponding metadata file into a structured archive package, and upload the structured archive package to the hospital information system.

[0068] Specifically, a pre-defined event list can be read from the final global context object. Based on the start timestamp and estimated end time of each event in the list, the anonymized video stream is divided into multiple video segments along the timeline. Each video segment uniquely corresponds to a key neonatal event, achieving event-based classification of image data. This solves the problems of traditional neonatal image data lacking event markers and being poorly managed, thus enhancing the clinical reusability of image data. Subsequently, a metadata file can be generated for each video segment. This metadata file contains information such as the neonatal identity, event type, and anonymization processing details, achieving a standardized description of the video segment. Furthermore, embedded dynamic lineage tags can be extracted from the video segments and compared with the dynamic lineage tags in the final global context object to verify whether identity information tampering or image frame replacement occurred during the anonymization and segmentation process, further ensuring the consistency between the image data and the neonatal identity. If the tag verification results are consistent, it means that the image data has not been tampered with and the identity association is correct. Then, each video segment and the corresponding metadata file can be packaged into a structured archive. This archive has the characteristics of identity binding, event marking, privacy desensitization, and complete metadata. It is then uploaded to the Baby Cloud Hospital Information System to realize the standardization and structured archiving of newborn image data, effectively improving the management efficiency and traceability of newborn image data.

[0069] The aforementioned method first identifies newborns through wristband near-field communication, acquires an initial video stream, and generates dynamic bloodline tag embedded frame images, solving the problem of lack of source binding between images and newborn identity, and achieving a strong correlation between identity information and image data. Secondly, it extracts features from multimodal sensor data and combines them with temporal convolutional networks and obstetric knowledge graph reasoning to generate acquisition trigger commands to control the start and stop of image acquisition, optimizing the global context object and solving the problems of missed acquisitions due to manual triggering and misjudgments due to simple sensor triggering, thus improving the accuracy of capturing key events in newborns. Furthermore, it identifies non-newborn sensitive areas through Mask R-CNN, dynamically adjusts the desensitization strategy based on event type, and uses a chaotic pixel replacement algorithm to achieve irreversible masking, generating a desensitized video stream and a final global context object, solving the problem of inaccurate privacy protection in obstetric scenarios and improving the targeting and security of privacy desensitization. Finally, it segments the desensitized video stream according to event information and generates metadata files, which can be packaged into structured archives after dynamic bloodline tag verification and uploaded to the hospital system, realizing structured archiving of newborn image data, improving the clinical reuse value and management efficiency of the data, and ensuring the accuracy of data archiving.

[0070] In one embodiment, a dynamic lineage tag is generated based on the UID and the binding timestamp, and the dynamic lineage tag is embedded into the initial image of each frame of the initial video stream through frequency domain transformation to obtain a lineage-bound video stream, including:

[0071] A hash algorithm is used to calculate the bound timestamp to obtain a 256-bit dynamic factor; the UID and the dynamic factor are then XORed bitwise to generate a 128-bit dynamic lineage tag.

[0072] The luminance component of each frame of the initial video stream is divided into 8×8 blocks, and a discrete cosine transform is performed on each block to obtain the frequency domain coefficient matrix of each block.

[0073] The dynamic lineage tag is converted into a binary bit stream. The intermediate frequency coefficients of each block are selected as the embedding carrier according to the Zigzag scanning order. By adjusting the numerical characteristics of the intermediate frequency coefficients, the corresponding bits of the binary bit stream are embedded into the intermediate frequency coefficients of each block one by one, and the embedded block frequency domain coefficient matrix is ​​obtained.

[0074] Perform inverse discrete cosine transform on each embedded block frequency domain coefficient matrix to obtain the pixel matrix of each block;

[0075] The pixel matrices of each block are spliced ​​and integrated according to the original image block positions during the block segmentation process to obtain the image frame with embedded tags;

[0076] The image frames after each embedded tag are combined sequentially in chronological order to obtain the bloodline-bound video stream.

[0077] Specifically, the SHA-256 hash algorithm can be used to calculate the bound timestamp. This algorithm is collision-resistant and one-way, ensuring that the calculation result of the bound timestamp is unique and cannot be reversed, effectively avoiding the risk of dynamic factors being tampered with or the bound timestamp being restored in reverse. For example, the bound timestamp is first converted into a continuous binary bit stream, and then this binary bit stream is input into the SHA-256 hash algorithm module. After hashing, a 256-bit hash value is output. Subsequently, the 256-bit hash value is normalized by modulo operation to obtain a 256-bit dynamic factor with a value range of [0,1). This process utilizes the dynamic nature of the bound timestamp to give the dynamic factor the characteristics of uniqueness and real-time updates, avoiding the tag predictability problem caused by using a fixed factor, and laying the foundation for the uniqueness of subsequent dynamic lineage tags. After generating the dynamic factor, the newborn's wristband UID is first converted into a 128-bit binary stream. Then, this 128-bit binary stream is XORed with the first 128 bits of the 256-bit dynamic factor to obtain a 128-bit dynamic kinship tag. The core function of the bitwise XOR operation is to combine the uniqueness of the UID with the dynamism of the dynamic factor, ensuring the uniqueness of each newborn's dynamic kinship tag while also ensuring the tag's timeliness by incorporating the time dimension of the bound timestamp. This effectively prevents tag duplication and identity theft. As an illustration, the dynamic kinship tag can be generated using the following formula:

[0078]

[0079] Here, Tag represents the generated 128-bit dynamic lineage tag. The 128-bit unique identifier binary stream representing the newborn's wristband. This represents the first 128 bits of the hash value extracted after performing a SHA-256 hash operation on the bound timestamp TS. This represents the bitwise XOR operator.

[0080] Furthermore, since the human eye is far less sensitive to changes in the luminance component than the chrominance component, dynamic lineage tags can be embedded in the luminance component of each initial frame of the initial video stream. This minimizes the impact of tag embedding on the visual effect of the image. Simultaneously, the luminance component contains the main structural information of the image, facilitating stable tag embedding and subsequent extraction. For example, a non-overlapping, row-by-row, column-by-column sliding block method can be used to perform 8×8 block processing on the luminance component image. During the block division, 8 pixels are used as a block unit. For instance, starting from the top-left corner (0,0) of the luminance component image, blocks can be slid horizontally sequentially. When a row of blocks is completed without exceeding the image width, 8 pixels are slid vertically downwards to continue dividing the blocks until the entire luminance component image is covered. During the block division process, the original position coordinates of each block are recorded simultaneously, including the top-left row and column index of the block, as well as the number of rows and columns in the block, to provide a basis for the accurate stitching of the subsequent block pixel matrix. Subsequently, a Discrete Cosine Transform (DCT) can be performed on each block. The DCT is a core mathematical technique for converting an image from the spatial domain to the frequency domain, transforming pixel values ​​into frequency coefficients. Low-frequency coefficients correspond to the overall brightness and large-scale structural features of the image, mid-frequency coefficients correspond to the detailed features, and high-frequency coefficients correspond to noise and edge details. In this embodiment, mid-frequency coefficients are selected as the tag embedding carrier, thus balancing the concealment and compression resistance of the tag embedding, avoiding significant visual distortion after tag embedding, and ensuring that the tag is not easily lost during image compression. The formula for calculating the DCT is as follows:

[0081]

[0082] in, This represents the frequency domain coefficients in the u-th row and v-th column of the frequency domain coefficient matrix divided into 8×8 blocks. and The values ​​range from 0 to 7, corresponding to the row and column indices of the frequency domain coefficients, respectively. and The normalized coefficients of the discrete cosine transform, when hour when hour , The rules for taking values ​​and exactly the same, This represents the pixel value in the x-th row and y-th column of an 8×8 block spatial domain image. and The values ​​range from 0 to 7, corresponding to the row and column coordinates of the pixels within the block, respectively.

[0083] Specifically, after calculating the block frequency domain coefficients, the 128-bit dynamic lineage tag can be converted into a continuous 128-bit binary bit stream in order from high to low bits. Then, the frequency domain coefficient matrix of each 8×8 block is traversed according to the Zigzag scanning order. The specific path of the Zigzag scan is to start from the upper left corner (0,0) of the frequency domain coefficient matrix, and traverse the coefficients sequentially along the Z-shaped trajectory. First, it moves horizontally from (0,0) to (0,7), then traverses downward along the diagonal to (7,0), and finally traverses to (7,7). In this embodiment, the mid-frequency coefficients of 2≤u≤5 and 2≤v≤5 in the Zigzag scanning order can be selected as the tag embedding carrier. The coefficients in this region do not involve the core brightness information of the image and are not easily lost by common image compression algorithms, which can effectively ensure the stability of tag embedding. By adjusting the numerical characteristics of the intermediate frequency (IF) coefficients, corresponding bits of the binary bitstream can be embedded one by one into the IF coefficients of each block. For example, a bit-plane replacement embedding method can be used. For each bit to be embedded, if the bit is 1, the value of the corresponding IF coefficient is increased by the embedding strength coefficient; if the bit is 0, the value of the corresponding IF coefficient is decreased by the embedding strength coefficient. The value of the embedding strength coefficient needs to be adaptively adjusted according to the brightness characteristics of the block image to ensure that the peak signal-to-noise ratio (PSNR) of the image after embedding the label is greater than 40 dB compared with the original image, avoiding distortion perceptible to the human eye. The adaptive adjustment of the embedding strength coefficient is based on the average brightness of the blocks. When the average brightness of the blocks is high, the embedding strength coefficient is appropriately increased to ensure the reliability of label extraction. In addition, when the average brightness of the blocks is low, the embedding strength coefficient can be decreased to avoid generating obvious noise and ensure the concealment and stability of the embedding operation. The frequency domain coefficient matrix after embedding is denoted as... .

[0084] After tag embedding is complete, inverse discrete cosine transform can be performed on each embedded block frequency domain coefficient matrix to convert the frequency domain coefficients back into pixel matrices in the spatial domain. Inverse discrete cosine transform is the inverse operation of discrete cosine transform, and its calculation formula is as follows:

[0085]

[0086] in, This represents the pixel value in the x-th row and y-th column within the block obtained after the inverse discrete cosine transform. These are the frequency domain coefficients after embedding the tags. , The definition is consistent with that in the discrete cosine transform. , The values ​​range from 0 to 7, corresponding to the row and column coordinates of pixels within the block. During the inverse transform, the spatial domain pixel information containing the dynamic lineage label is restored by reconstructing the frequency domain coefficients after embedding the label. Since only the intermediate frequency coefficients are adjusted, the reconstructed block pixel matrix can still maintain the structure and visual effect of the original image.

[0087] Specifically, based on the original position coordinates recorded during segmentation, the pixel matrix of each segment is placed back into its corresponding position in the entire frame image. First, all segments within the same row are horizontally stitched together. Then, the stitched rows are stacked vertically until all segments are stitched together, resulting in a tagged image frame. This tagged image frame is identical in size and structure to the original image frame, containing dynamic kinship tag information only in the mid-frequency details of the luminance component. Finally, the tagged image frames are sequentially combined in chronological order to obtain a kinship-bound video stream. The chronological order is based on the system timestamp corresponding to each image frame, ensuring the temporal continuity of the video stream matches the order of events. Each image frame contains a unique dynamic kinship tag, realizing the physical binding of newborn identification information to each frame. This bloodline-bound video stream not only retains the visual features and temporal information of the original video stream, but also possesses the characteristics of unalterable and traceable identity. Therefore, regardless of whether it is compressed, transmitted, or edited in subsequent processes, the identity of the newborn can be verified by extracting the dynamic bloodline tag in the image frame. This eliminates the problem of separation between image and identity and confusion of records from the source of data, and provides technical support for the standardized management of newborn image records.

[0088] In one embodiment, the pre-defined temporal convolutional network includes a feature extraction module, a residual module, a global average pooling layer, and a fully connected layer, wherein the residual module is constructed based on an attention gating mechanism. Figure 2 As shown, feature extraction is performed on the audio stream and the accelerometer sensor data stream to obtain a multimodal feature tensor. This multimodal feature tensor is then input into a pre-defined temporal convolutional network, which outputs an event probability vector, including:

[0089] S201: Perform frame segmentation on the audio stream. Based on the preset frame length and frame shift, apply a Hamming window to each frame of audio, calculate the first 13-dimensional Mel frequency cepstral coefficients, combine short-time energy and zero-crossing rate to obtain the single-frame audio feature vector, and stack the single-frame audio feature vectors of consecutive preset frames to obtain the audio feature tensor.

[0090] Specifically, framing can be performed based on the temporal characteristics of the audio signal. This means that the framing operation uses the audio sampling rate as a reference, dividing the continuous audio stream into non-overlapping and temporally continuous audio frames according to a preset frame length and frame shift. The frame shift can be set to 1 / 2 to 1 / 3 of the frame length to ensure the continuity of features between adjacent frames and avoid the loss of critical information. For each audio frame, pre-emphasis processing can be performed first, using a first-order high-pass filter to enhance the high-frequency components of the audio signal and compensate for the attenuation of high-frequency signals during audio transmission. Then, a Hamming window is applied to reduce spectral leakage caused by framing. The expression for the Hamming window is: ,in This represents the weight value of the nth sampling point in the Hamming window. For the index of sampling points within the window, This represents the total number of sampling points in a single frame of audio.

[0091] After applying the window function, a Fast Fourier Transform (FFT) can be performed on each frame of audio to convert the time-domain signal into a frequency-domain signal, obtaining the frequency-domain amplitude spectrum. Based on the human auditory system's perception characteristics of different frequencies, a Mel filter bank can be constructed to filter the frequency-domain amplitude spectrum, converting the filtered result into logarithmic energy. Performing a Discrete Cosine Transform (DCT) on the logarithmic energy allows the extraction of the first 13 dimensions of coefficients as Mel frequency cepstral coefficients. These coefficients effectively characterize the acoustic features of the audio, fitting the audio characteristics of events such as newborn crying. Furthermore, the short-time energy and zero-crossing rate of each frame of audio can be calculated. The short-time energy can be obtained by calculating the sum of squares of all sampling points within the frame, reflecting the intensity changes of the audio signal. The zero-crossing rate can be obtained by counting the number of sign changes at sampling points within the frame, distinguishing between voiced and unvoiced sounds. Concatenating the 13-dimensional Mel frequency cepstral coefficients, short-time energy, and zero-crossing rate of a single frame of audio according to their dimensions forms the audio feature vector of that frame. By stacking the single-frame audio feature vectors of consecutive preset frames along the time dimension in chronological order, a three-dimensional audio feature tensor can be constructed. The dimensions of this tensor correspond to the number of frames, the feature dimension, and the number of channels, respectively, which can preserve complete temporal information.

[0092] S202: Based on the preset window length and step size, a sliding window is used to process the acceleration sensor data stream. The mean, variance, peak-to-peak value, energy entropy and spectral peak features are extracted from the triaxial acceleration data in each sliding window to generate the acceleration feature vector corresponding to each sliding window. The acceleration feature vectors of multiple consecutive sliding windows are stacked to obtain the acceleration feature tensor.

[0093] Specifically, for the data stream acquired by the triaxial accelerometer, a sliding window method can be used for feature extraction. The length and step size of the sliding window can be set based on the sampling frequency of the accelerometer. The window length needs to cover the smallest time unit of events such as newborn limb movements and touch, while the step size can be set to half the window length to balance the timeliness and redundancy of feature extraction. Feature calculations are performed on the X, Y, and Z axis acceleration data within each sliding window. The mean can be obtained by the arithmetic mean of all sampling points within the window, reflecting the average intensity of acceleration within that time period. The variance can be obtained by calculating the average of the squares of the differences between the sampling points and the mean, characterizing the dispersion of the acceleration data. The peak-to-peak value is the difference between the maximum and minimum acceleration values ​​within the window, used to capture the intensity of sudden movements. The energy entropy can be obtained by calculating the power spectral density of the acceleration signal, converting it into a probability distribution, and then calculating it according to the definition of entropy, reflecting the complexity of the energy distribution of the acceleration signal. The spectral peak value can be obtained by performing a fast Fourier transform on the acceleration signal and extracting the maximum value in the frequency domain, used to capture the main frequency components of the movement.

[0094] By concatenating the mean, variance, peak-to-peak value, energy entropy, and spectral peak values ​​extracted from the three-axis acceleration data according to their dimensions, an acceleration feature vector corresponding to each sliding window can be formed. The vector dimension is the sum of the number of features in each of the three axes. Stacking the acceleration feature vectors of multiple consecutive sliding windows along the time dimension in temporal order can construct a three-dimensional acceleration feature tensor. The dimensions of this tensor correspond to the number of windows, the feature dimension, and the number of channels, respectively, achieving a precise characterization of the temporal features of neonatal motion.

[0095] S203: Time-align the audio feature tensor and the acceleration feature tensor to obtain a multimodal feature tensor; input the multimodal feature tensor into a preset temporal convolutional network, extract temporal features from the multimodal feature tensor through the feature extraction module to obtain initial temporal features, calculate the gate coefficients through the residual module, and perform weighted enhancement on the initial temporal features based on the gate coefficients to obtain weighted temporal features.

[0096] Specifically, due to the difference in sampling frequencies between the audio stream and the accelerometer data stream, time alignment is required for the audio feature tensor and the acceleration feature tensor to ensure temporal consistency of multimodal features. Time alignment is based on a unified system timestamp, and the time dimension of the acceleration feature tensor is adjusted using linear interpolation to ensure a one-to-one correspondence between the time nodes of the acceleration feature tensor and the frame time nodes of the audio feature tensor. Subsequently, the aligned audio feature tensor and the acceleration feature tensor are concatenated along the feature dimension to form a multimodal feature tensor. This tensor simultaneously contains acoustic and motion features, comprehensively reflecting the newborn's state changes.

[0097] The multimodal feature tensor is input into a pre-defined temporal convolutional network. The feature extraction module of this network consists of multiple one-dimensional convolutional layers. The convolutional kernels slide along the time dimension, capturing local temporal features of different durations through convolutional kernels of different sizes. The expression for the convolution operation is as follows: ,in The output features of the I-th convolutional layer at time t are... The weights are the kernel weights of the I-th convolutional layer, and * represents the one-dimensional convolution operator. The input features are those of the (I-1)th layer. For the bias term of layer 1, It is a ReLU activation function, used to introduce non-linear feature representation capabilities.

[0098] The initial temporal features output by the feature extraction module can then be input into the residual module based on the attention gating mechanism. This module can adaptively weight the initial temporal features by calculating gating coefficients, where the expression for calculating the gating coefficients is as follows: ,in For time The gating coefficient has a value range of [0,1]. For the gated weight matrix, The gating bias term is denoted by , and the sigmoid function is used as the activation function. By multiplying the gating coefficients element-wise with the initial temporal features, the strong response features related to the event are weighted and strengthened, while the weak response features of noise are suppressed, resulting in weighted temporal features. This process effectively improves the network's ability to capture key event features and enhances the discriminative power of the features.

[0099] S204: The weighted temporal features are reduced in dimensionality by the global average pooling layer of the pre-defined temporal convolutional network. The reduced features are then input into the fully connected layer for classification operations. The output is an event probability vector containing the probabilities of crying, touching, opening eyes, and no events.

[0100] Specifically, the weighted temporal features are input into a global average pooling layer. This layer averages the temporal dimension of the features, converting the variable-length features into fixed-dimensional feature vectors. This preserves the core information of the temporal features while significantly reducing the number of network parameters and mitigating the risk of overfitting. The dimensionality-reduced feature vectors are then input into a fully connected layer, which may contain two hidden layers. The first hidden layer uses the ReLU activation function for non-linear transformation, and the second hidden layer uses the softmax activation function to map the feature vectors into a probability distribution, outputting an event probability vector. The expression for the softmax function is as follows: ,in Let be the probability of the i-th type of event. This represents the raw score of the i-th type of event output by the fully connected layer. This is the event category index, with values ​​from 1 to 4, corresponding to crying events, touching events, eye-opening events, and no events, respectively. In the final output event probability vector, each element has a value in the range [0,1], and the sum of all elements is 1, intuitively reflecting the probability of occurrence of each type of event.

[0101] In one embodiment, rule-based reasoning is performed by querying the obstetric knowledge graph based on the event probability vector. When the reasoning result confirms that an event has occurred, event information is generated, including:

[0102] Obtain a preset first threshold and a second threshold. When the probability of any event in the event probability vector exceeds the first threshold, the event whose probability exceeds the first threshold is identified as a candidate event.

[0103] Extract the timestamp, audio energy parameter, and peak acceleration parameter corresponding to the candidate event as context attributes, and construct a query graph based on the context attributes;

[0104] Input the query graph into the preset obstetrics knowledge graph for matching query to obtain entity association matching results; the obstetrics knowledge graph includes entities, attributes and preset medical common sense defined in the form of rules. Entities include newborns, various target events and image acquisition equipment. Attributes include event duration and event association relationship. Medical common sense includes event duration rules and event association verification rules.

[0105] By using the rule-based reasoning engine of the obstetrics knowledge graph and combining entity association matching results, we can verify whether candidate events conform to preset medical common sense and output the event confidence adjustment factor.

[0106] The event confidence adjustment factor is multiplied by the event probability corresponding to the candidate event in the event probability vector to obtain the final confidence and generate the corresponding inference result.

[0107] If the final confidence level exceeds the second threshold, the inference result is that the event has occurred. The event type is determined according to the type of the candidate event. The center time of the feature extraction window corresponding to the candidate event is used as the start timestamp. The estimated end time is determined based on the event duration rules in the obstetrics knowledge graph. The feature extraction window is the audio frame window corresponding to the Hamming window or the acceleration feature extraction window corresponding to the sliding window.

[0108] Integrate event type, start timestamp, and estimated end time to generate corresponding event information.

[0109] Specifically, both the first and second thresholds can be determined based on statistical analysis of a large sample of obstetric neonatal events. The first threshold is used to initially screen potentially valid events, balancing the event recall rate and false positive rate, avoiding missed detections of critical events due to an excessively high threshold, or the introduction of too many noisy events due to an excessively low threshold. The second threshold is used to finally confirm the authenticity of the event, and its value can be higher than the first threshold to ensure that only high-confidence events are included in the subsequent processing. Illustratively, each element in the event probability vector can be iterated and judged. Each element in the event probability vector corresponds to the probability of occurrence of crying, touching, eye-opening, and no event, respectively. When the value of any element exceeds the first threshold, it indicates that the event corresponding to that element has a certain probability of occurrence and can be identified as a candidate event. The initial event probability corresponding to the candidate event can be recorded simultaneously to provide a basis for subsequent confidence calculations.

[0110] Furthermore, contextual attributes corresponding to candidate events can be extracted, such as timestamps, audio energy parameters, and acceleration peak parameters. The timestamp is the center time of the feature extraction window corresponding to the candidate event, i.e., the center time of the Hamming window corresponding to the audio frame or the center time of the sliding window corresponding to the acceleration feature extraction, used to accurately locate the time node of the event occurrence; the audio energy parameter is the average short-time energy of a single frame of audio within the feature extraction window, reflecting the acoustic intensity characteristics at the time of the event; the acceleration peak parameter is the maximum value of the peak-to-peak values ​​of the three-axis acceleration within the feature extraction window, characterizing the motion intensity characteristics at the time of the event. Subsequently, with candidate events as core nodes and timestamps, audio energy parameters, and acceleration peak parameters as attribute nodes, semantic edges can be used to establish the association between core nodes and attribute nodes. The type of semantic edge is defined as "possessing," meaning that candidate event nodes are connected to each attribute node through "possessing" edges. Ultimately, a structured query graph can be formed, ensuring that the query graph semantically aligns with the entity and attribute system of the obstetrics knowledge graph.

[0111] Specifically, the obstetrics knowledge graph is a structured semantic network built upon obstetric clinical medical literature, clinical practice data, and expert experience. It is stored in RDF (Resource Description Framework) format and supports semantic retrieval using the SPARQL query language. Entities in this knowledge graph can include three core types: newborn entities, various target event entities (including crying events, touch events, eye-opening events, etc.), and image acquisition device entities (representing the data acquisition source). Attributes in the graph can include event duration and event relationships. The event duration attribute defines the typical duration range of various target events, such as the typical duration of crying events and the typical duration of touch events. The event relationship attribute defines the association characteristics between events and audio and acceleration parameters, such as the audio energy range corresponding to crying events and the peak acceleration range corresponding to touch events. Preset medical common sense defined in the form of rules can include event duration rules and event relationship verification rules. Event duration rules clarify the reasonable time intervals for various events, while event relationship verification rules clarify the logical association between events and various parameters. For example, a crying event must satisfy the condition that the audio energy parameter is higher than a certain threshold and the peak acceleration parameter is within a specific range.

[0112] The constructed query graph is input into the knowledge graph for matching queries. During the matching query process, the candidate event nodes in the query graph are first semantically matched with the target event entities in the knowledge graph to determine the knowledge graph entity type corresponding to the candidate event. Then, the attribute nodes in the query graph are associated with the attributes corresponding to the entity type to determine whether the context attributes of the candidate event intersect with the attributes of the event entity in the knowledge graph. Finally, the entity association matching result is output, which includes information such as the matching degree between the candidate event and the knowledge graph entity, and the degree of conformity between each context attribute and the corresponding entity attribute. Subsequently, the rule reasoning engine of the obstetric knowledge graph can be used to perform reasoning verification. This rule reasoning engine can be a production rule engine, which converts the medical common sense rules in the knowledge graph into production rules in the form of "IF-THEN" (e.g., IF event is crying AND audio energy parameters are in the range [E1,E2] AND peak acceleration parameters are in the range [A1,A2] AND duration is in the range [T1,T2] THEN event verification passed). During the reasoning process, the rule-based reasoning engine reads the entity association matching results and substitutes the contextual attributes of candidate events into the corresponding production rules for logical judgment, verifying one by one whether the candidate events conform to the event duration rules and event association verification rules. For example, if the candidate event is a crying event, it needs to verify whether its corresponding audio energy parameters are within the range of audio energy attributes of crying events in the knowledge graph, whether the peak acceleration parameters meet the motion characteristic requirements of crying events, and whether the event duration estimated based on the feature extraction window duration is within the normal duration range of crying events. Based on the verification results, an event confidence adjustment factor can be output. This adjustment factor is used to quantify the degree to which the candidate event conforms to medical common sense, and its calculation formula is:

[0113]

[0114] in, This represents the event confidence adjustment factor, with a value range of [0,1]. , , The weighting coefficients are for the compliance of event duration, audio parameters, and acceleration parameters, respectively. The sum of the three is 1. The weighting coefficients are determined by expert scoring based on the importance of each verification dimension to the event judgment. The event duration conformity is calculated as the overlap ratio between the estimated duration of the candidate event and the normal duration range of the event in the knowledge graph, with a range of [0,1]. The audio parameter conformity is calculated as the degree of fit between the audio energy parameters of the candidate event and the range of the audio energy attributes of the event in the knowledge graph, with a range of [0,1]. The acceleration parameter conformity is calculated as the degree of fit between the peak acceleration parameter of the candidate event and the range of the peak acceleration attribute of the event in the knowledge graph, with a range of [0,1].

[0115] Specifically, multiplying the event confidence adjustment factor by the initial event probability corresponding to the candidate event in the event probability vector allows for the organic integration of model prediction results and medical common sense verification results. This ensures that the final confidence level reflects both the probability of the event occurring at the data feature level and the rationality at the clinical common sense level. If the final confidence level exceeds the second threshold, it indicates that the candidate event has high credibility at both the data feature and medical common sense levels, and the inference result can be considered a confirmed event occurrence. At this point, the event type in the event information can be directly determined based on the type of the candidate event, ensuring the accuracy of the event type. The center time of the feature extraction window corresponding to the candidate event can be used as the starting timestamp, which accurately corresponds to the moment when the event features are most significant, ensuring the accuracy of the event start time record. In addition, the estimated end time can be determined based on the event duration rules in the obstetrics knowledge graph. For example, the median of the normal duration range of the event in the knowledge graph can be selected and added to the starting timestamp to obtain the estimated end time. This method not only conforms to the general duration pattern of clinical events but also provides a reasonable time reference for subsequent video acquisition start / stop control and video segmentation.

[0116] Finally, the determined event types, start timestamps, and estimated end times are structurally integrated to generate corresponding event information. This event information is stored in a standardized data format, including fields such as a unique event identifier, event type label, start timestamp string, and estimated end timestamp string. This ensures the standardization and scalability of the event information, allowing it to be directly accessed by subsequent video segmentation and desensitization strategy adjustment modules, and also enabling efficient association with other data in the global context object. This provides data support for the event-based management of neonatal imaging data.

[0117] In one embodiment, the preset Mask R-CNN instance segmentation network includes a region proposal network and a classification branch. The region proposal network includes 16×16 anchor boxes and 32×32 anchor boxes. The frame image is input into the preset Mask R-CNN instance segmentation network to identify sensitive regions that are not newborn subjects. The desensitization strategy is dynamically adjusted based on the event type in the optimized global context object. For sensitive regions, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. A chaotic pixel replacement algorithm is used to irreversibly mask the sensitive regions, generating desensitized video frames. The sensitive region categories and processing parameters during the desensitization process are recorded, generating desensitization operation metadata, including:

[0118] The frame image is input into the preset Mask R-CNN instance segmentation network. Multiple candidate region boxes are generated through the 16×16 and 32×32 anchor boxes of the region proposal network. The region of interest pooling operation is performed on the candidate region boxes to obtain the feature map corresponding to each candidate region box.

[0119] The feature maps are input into the classification branch of the preset Mask R-CNN instance segmentation network. The classification branch determines the category of the target corresponding to each feature map and predicts the segmentation mask and confidence level of each target. The bounding box, category label, segmentation mask and confidence level of each target are output. Targets with confidence levels exceeding the preset confidence threshold are identified as detection targets. Targets with category labels of medical staff faces, department logos or background infants are selected as sensitive regions from the detection targets.

[0120] The event type is extracted from the optimized global context object, and the desensitization strategy is dynamically adjusted according to the event type to obtain the target desensitization strategy; the UID in the optimized global context object is concatenated with the current frame timestamp to obtain the concatenation result, and a hash operation is performed on the concatenation result to obtain the initial seed of the chaotic mapping;

[0121] For each sensitive region, the initial seed of the chaotic mapping is iteratively calculated based on the Logistic chaotic mapping formula to generate a chaotic sequence equal to the total number of pixels in the sensitive region, and a permutation index array equal to the total number of pixels in the sensitive region is generated based on the chaotic sequence.

[0122] The pixels in the sensitive area are expanded into a one-dimensional pixel array in row order. The pixel positions of the one-dimensional pixel array are rearranged according to the permutation index array to obtain the rearranged one-dimensional pixel array.

[0123] The rearranged one-dimensional pixel array is filled back into the sensitive area according to the original area coordinates. The filled sensitive area is then merged with the unprocessed area in the frame image to generate a desensitized video frame.

[0124] Record the category of the sensitive region, the bounding box coordinates, the initial seed of the chaotic mapping, and the generation parameters of the permutation index array to obtain the metadata of the desensitization operation.

[0125] Specifically, the frame image is input into a pre-defined Mask R-CNN instance segmentation network. Its region proposal network is configured with anchor boxes of two sizes: 16×16 and 32×32. The former is suitable for small-scale sensitive targets such as department logos, while the latter is suitable for slightly larger-scale sensitive targets such as faces of medical staff and infants in the background. Multi-scale anchor boxes can cover non-newborn main targets of different sizes, improving the recall and accuracy of sensitive region detection. Illustratively, the region proposal network can slide on the feature map of the frame image based on the anchor boxes, generating multiple candidate region boxes containing potential targets. The number of candidate region boxes is determined by the number of anchor boxes and the image size. Each candidate region box corresponds to a set of coordinate information to represent the spatial range of the potential target. Region pooling is performed on all candidate region boxes, mapping candidate region boxes of different sizes to feature maps of a fixed size. Max pooling preserves key features within the candidate regions and eliminates the impact of candidate region box size differences on subsequent processing, ensuring that the feature map corresponding to each candidate region box has consistent dimensions, providing a standardized format for the input of the classification branch.

[0126] Specifically, the feature map after pooling the region of interest is input into the classification branch of the pre-defined Mask R-CNN instance segmentation network. This branch, composed of convolutional and fully connected layers, is primarily responsible for three tasks: category determination, segmentation mask prediction, and confidence assessment. Category determination distinguishes the target category corresponding to the feature map, and the output category labels can include five categories: newborn subject, medical staff face, department logo, background infant, and other backgrounds. Segmentation mask prediction generates pixel-level target region masks. The mask is a binary image with the same size as the feature map; regions with a mask value of 1 correspond to the precise pixel range of the target. Compared to bounding boxes, this more accurately defines the boundaries of sensitive regions, preventing the desensitization operation from spreading to the newborn subject region. Confidence assessment quantifies the reliability of the detection results, outputting a confidence score in the range [0,1]. A higher score indicates a more reliable target detection result. The classification branch ultimately outputs the bounding box coordinates, category label, segmentation mask, and confidence score for each target. Subsequently, the detection results are filtered based on a pre-set confidence threshold, retaining only targets with confidence scores exceeding the threshold as detection targets, further eliminating noisy detection results. By selecting targets from the detection targets that are categorized as medical staff faces, department logos, or background infants, these can be identified as sensitive areas. This ensures that desensitization is performed only on privacy-related areas that are not the newborn subject, thus fully preserving the image information of the newborn subject and guaranteeing the clinical needs for using newborn images.

[0127] Furthermore, the privacy exposure risk of sensitive areas varies under different event types. For example, in crying events, the probability of medical staff's faces being exposed is higher, requiring stronger desensitization in such areas. In touch events, background infants are easily confused with the target newborn, necessitating higher desensitization priority for background infant areas. Therefore, event types can be extracted from the optimized global context object, and desensitization strategies can be dynamically adjusted based on the correlation between event types and sensitive area categories. For example, by adjusting the sensitive area category weights and chaotic mapping-related parameters, a target desensitization strategy can be obtained, providing differentiated configurations for subsequent chaotic sequence generation. Concatenating the UID in the optimized global context object with the current frame timestamp involves sequentially connecting the binary bit stream of the UID and the binary bit stream of the current frame timestamp from high to low bits, resulting in a continuous concatenation result. A hash algorithm with excellent collision resistance and one-wayness is selected. The hash operation is performed on the splicing result, and the result is normalized to the interval [0,1) to obtain the initial seed of the chaotic mapping. This seed integrates the uniqueness of the newborn's identity and the temporal dynamics of the frame image, ensuring that the chaotic sequence corresponding to each frame image and each sensitive region has uniqueness and unpredictability.

[0128] Specifically, for each sensitive region, the initial seed of the chaotic mapping can be iteratively calculated based on the improved piecewise Logistic chaotic mapping formula to generate a chaotic sequence equal to the total number of pixels within the sensitive region. The mathematical expression of this formula can be:

[0129]

[0130]

[0131]

[0132] in, The initial seed for the chaotic mapping is obtained by summing and normalizing the bits of the UID and the current frame timestamp after bitwise XOR. For the UID Bit binary value, The range of values ​​is to ; The timestamp of the current frame Bit binary value, The range of values ​​is to ; This is a bitwise XOR operation used to combine the binary information of UID and timestamp, improving the randomness of the initial seed; For the first The chaotic sequence value of the next iteration has a range of values. ; and The piecewise setting of the control parameters for chaos avoids the defect of the traditional Logistic mapping getting stuck in a periodic state in some intervals. and This is the category weight adjustment coefficient for sensitive areas, with a value range of [value range missing]. This is used to adjust the degree of influence of the category weight of sensitive regions on chaotic sequences; The sensitive area category weight is determined by the target desensitization strategy based on the event type and the sensitive area category, and is used to achieve differentiated desensitization intensity for different sensitive areas; For the first The permutation index of pixels, with values ​​ranging from . to ; This represents the total number of pixels within the sensitive area. For floor operations, For modulo operations, ensure that the permutation index always falls within the range of the total number of pixels in the sensitive region. Based on the initial seed. The iteration begins, with the number of iterations equal to the total number of pixels in the sensitive region. Each iteration yields a chaotic sequence value. By combining all the chaotic sequence values ​​obtained from the iterations in order, a chaotic sequence with the same length as the total number of pixels in the sensitive region can be formed. Then, based on the chaotic sequence values ​​and the above permutation index formula, a permutation index array with the same length as the total number of pixels in the sensitive region can be generated.

[0133] Specifically, the pixels within the sensitive area are unfolded into a one-dimensional pixel array in row-wise order. Starting from the top-left pixel of the sensitive area, each row of pixels is read sequentially along the horizontal direction until all pixels within the sensitive area are read, forming a linearly arranged one-dimensional pixel array. Then, the pixel positions in the one-dimensional pixel array are rearranged according to a permutation index array. Specifically, the pixel at position m in the one-dimensional pixel array is moved to the position corresponding to the m-th element in the permutation index array. After all pixel positions are permuted, a rearranged one-dimensional pixel array is obtained. This permutation process is irreversible, relying solely on the chaotic sequence and the permutation index array. It is impossible to reverse the original pixel order from the rearranged pixel array, effectively improving the security of sensitive area privacy protection. The rearranged one-dimensional pixel array is then filled back into the frame image according to the coordinates of the original sensitive area. The filling order is the same as the unfolding order, starting from the top-left corner of the sensitive area and filling the corresponding positions row-wise. After filling, the sensitive area is merged with the unprocessed newborn subject area in the frame image to generate a desensitized video frame. Furthermore, key parameters during the desensitization process can be recorded to generate desensitization operation metadata, such as category labels for sensitive regions, bounding box coordinates, initial seeds for chaotic mapping, and generation parameters for the permutation index array. The sensitive region category labels identify the type of the desensitized object, the bounding box coordinates locate the sensitive region's position in the frame image, the initial seed for chaotic mapping verifies the legality of the desensitization process, and the generation parameters for the permutation index array can include piecewise chaos control parameters, sensitive region category weight adjustment coefficients, and sensitive region category weights, used to reproduce the desensitization operation process. Appending the desensitization operation metadata to the optimized global context object enables unified management of desensitized information, newborn identity information, and event information, providing data support for subsequent image data verification and traceability.

[0134] Based on the same inventive concept, this application also provides an AI-based neonatal image recording and detection system for implementing the AI-based neonatal image recording and detection method described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more AI-based neonatal image recording and detection system embodiments provided below can be found in the limitations of the AI-based neonatal image recording and detection method described above, and will not be repeated here.

[0135] In one exemplary embodiment, such as Figure 3 As shown, an artificial intelligence-based neonatal image recording and detection system 300 is provided, including:

[0136] The identity binding and video acquisition module 301 is used to read the UID and newborn basic information in the wristband when the near-field communication signal of the wristband is detected, acquire the initial video stream in real time through the image acquisition device, obtain the current system timestamp as the binding timestamp, generate a dynamic bloodline tag based on the UID and the binding timestamp, and embed the dynamic bloodline tag into the initial image of each frame of the initial video stream through frequency domain transformation to obtain the bloodline binding video stream. The UID, binding timestamp, dynamic bloodline tag and newborn basic information are packaged to generate a global context object.

[0137] The intelligent event triggering module 302 is used to acquire audio streams and accelerometer data streams in real time, extract features from the audio streams and accelerometer data streams to obtain multimodal feature tensors, input the multimodal feature tensors into a preset temporal convolutional network, and output event probability vectors. Based on the event probability vectors, it queries the obstetrics knowledge graph to perform rule reasoning. When the reasoning result confirms the occurrence of an event, it generates event information, generates acquisition trigger instructions based on the event information, stores the event information in a preset event list, and appends the preset event list to a global context object to obtain an optimized global context object. The acquisition trigger instructions are used to control the start and stop of the acquisition of the initial video stream by the image acquisition device. The event information includes the event type, start timestamp, and estimated end time.

[0138] The privacy protection and desensitization module 303 is used to obtain the timestamp corresponding to each frame of the bloodline-bound video stream as the current frame timestamp, input the frame image into a preset Mask R-CNN instance segmentation network to identify sensitive areas of non-newborn subjects; dynamically adjust the desensitization strategy according to the event type in the optimized global context object; for sensitive areas, generate a chaotic sequence based on the UID and the current frame timestamp in the optimized global context object, irreversibly cover the sensitive areas through a chaotic pixel replacement algorithm, generate desensitized video frames, record the sensitive area category and processing parameters during the desensitization process, generate desensitization operation metadata, append the desensitization operation metadata to the optimized global context object to obtain the final global context object, and combine each desensitized video frame to obtain the desensitized video stream;

[0139] The file structuring and archiving module 304 is used to read a preset event list from the final global context object, and divide the desensitized video stream into multiple video segments along the timeline according to the start timestamp and estimated end time of each event in the preset event list; generate a metadata file for each video segment, and extract the embedded dynamic lineage tags from the video segments; compare and verify the embedded dynamic lineage tags with the dynamic lineage tags in the final global context object to obtain the tag verification result; if the tag verification result is consistent, package each video segment and the corresponding metadata file into a structured file package, and upload the structured file package to the hospital information system.

[0140] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the artificial intelligence-based neonatal image recording and detection method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0141] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the artificial intelligence-based neonatal image recording and detection method of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0142] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for detecting newborn image recordings based on artificial intelligence, characterized in that, The method includes: When a near-field communication signal from the wristband is detected, the UID and basic information of the newborn are read from the wristband. An initial video stream is acquired in real time through an image acquisition device. The current system timestamp is obtained as a binding timestamp. A dynamic bloodline tag is generated based on the UID and the binding timestamp. The dynamic bloodline tag is embedded into the initial image of each frame of the initial video stream through frequency domain transformation to obtain a bloodline-bound video stream. The UID, the binding timestamp, the dynamic bloodline tag, and the basic information of the newborn are packaged to generate a global context object. The system acquires audio streams and accelerometer data streams in real time, extracts features from these streams to obtain multimodal feature tensors, inputs these tensors into a preset temporal convolutional network to output event probability vectors, queries an obstetrics knowledge graph based on these event probability vectors for rule-based reasoning, generates event information when the reasoning result confirms the occurrence of an event, generates a collection trigger command based on the event information, stores the event information in a preset event list, and appends the preset event list to a global context object to obtain an optimized global context object. The collection trigger command controls the start and stop of the image acquisition device's acquisition of the initial video stream. The event information includes event type, start timestamp, and estimated end time. For each frame of the bloodline-bound video stream, the timestamp corresponding to the frame image is obtained as the current frame timestamp. The frame image is input into a preset Mask R-CNN instance segmentation network to identify sensitive regions that are not newborn subjects. The desensitization strategy is dynamically adjusted according to the event type in the optimized global context object. For the sensitive regions, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. The sensitive regions are irreversibly covered by a chaotic pixel replacement algorithm to generate desensitized video frames. The sensitive region categories and processing parameters during the desensitization process are recorded to generate desensitization operation metadata. The desensitization operation metadata is appended to the optimized global context object to obtain the final global context object. Combined with each of the desensitized video frames, a desensitized video stream is obtained. The preset event list is read from the final global context object. Based on the start timestamp and estimated end time of each event in the preset event list, the desensitized video stream is divided into multiple video segments along the time axis. A metadata file is generated for each video segment, and embedded dynamic lineage tags are extracted from the video segments. The embedded dynamic lineage tags are compared and verified with the dynamic lineage tags in the final global context object to obtain a tag verification result. If the tag verification result is consistent, each video segment and the corresponding metadata file are packaged into a structured archive package, and the structured archive package is uploaded to the hospital information system.

2. The method according to claim 1, characterized in that, The step of generating a dynamic lineage tag based on the UID and the binding timestamp, and embedding the dynamic lineage tag into the initial image of each frame of the initial video stream through frequency domain transformation to obtain a lineage-bound video stream, includes: A hash algorithm is used to calculate the bound timestamp to obtain a 256-bit dynamic factor; the UID and the dynamic factor are then XORed bitwise to generate a 128-bit dynamic lineage tag. The luminance component of each frame of the initial image in the initial video stream is divided into 8×8 blocks, and a discrete cosine transform is performed on each block to obtain the frequency domain coefficient matrix of each block. The dynamic lineage tag is converted into a binary bit stream. The intermediate frequency coefficients of each block are selected as embedding carriers according to the Zigzag scanning order. By adjusting the numerical characteristics of the intermediate frequency coefficients, the corresponding bits of the binary bit stream are embedded one by one into the intermediate frequency coefficients of each block to obtain the embedded block frequency domain coefficient matrix. Perform inverse discrete cosine transform on each of the embedded block frequency domain coefficient matrices to obtain the pixel matrix of each block; The pixel matrices of each block are spliced ​​and integrated according to the original image block positions during the block segmentation process to obtain the image frame with embedded tags; The image frames with embedded tags are combined sequentially in chronological order to obtain the bloodline-bound video stream.

3. The method according to claim 1, characterized in that, The preset temporal convolutional network includes a feature extraction module, a residual module, a global average pooling layer, and a fully connected layer, wherein the residual module is constructed based on an attention gating mechanism; the accelerometer data stream is triaxial acceleration data. The audio stream and the accelerometer data stream are subjected to feature extraction to obtain a multimodal feature tensor; The multimodal feature tensor is input into a preset temporal convolutional network, which outputs an event probability vector, including: The audio stream is segmented into frames. Based on the preset frame length and frame shift, a Hamming window is applied to each frame of audio. The first 13-dimensional Mel frequency cepstral coefficients are calculated. The single-frame audio feature vector is obtained by combining short-time energy and zero-crossing rate. The single-frame audio feature vectors of consecutive preset frames are stacked to obtain the audio feature tensor. Based on a preset window length and step size, a sliding window is used to process the acceleration sensor data stream. The mean, variance, peak-to-peak value, energy entropy, and spectral peak features are extracted from the triaxial acceleration data in each sliding window to generate an acceleration feature vector corresponding to each sliding window. The acceleration feature vectors of multiple consecutive sliding windows are stacked to obtain an acceleration feature tensor. The audio feature tensor and the acceleration feature tensor are time-aligned to obtain the multimodal feature tensor; The multimodal feature tensor is input into the preset temporal convolutional network. The feature extraction module extracts temporal features from the multimodal feature tensor to obtain initial temporal features. The residual module calculates the gate coefficients. The initial temporal features are then weighted and enhanced based on the gate coefficients to obtain weighted temporal features. The weighted temporal features are reduced in dimensionality by the global average pooling layer of the preset temporal convolutional network. The reduced features are then input into the fully connected layer for classification operations, and the output is the event probability vector containing the probabilities of crying events, touching events, eye-opening events, and no events.

4. The method according to claim 3, characterized in that, The rule-based reasoning based on the event probability vector query of the obstetric knowledge graph, when the reasoning result confirms the occurrence of the event, generates event information, including: Obtain a preset first threshold and a second threshold. When the probability of any event in the event probability vector exceeds the first threshold, determine the corresponding event whose probability exceeds the first threshold as a candidate event. Extract the timestamp, audio energy parameter, and peak acceleration parameter corresponding to the candidate event as context attributes, and construct a query graph based on the context attributes; The query graph is input into the preset obstetrics knowledge graph for matching query to obtain entity association matching results; wherein, the obstetrics knowledge graph includes entities, attributes and preset medical common sense defined in the form of rules, the entities include newborns, various target events and the image acquisition device, the attributes include event duration and event association relationship, and the medical common sense includes event duration rules and event association verification rules; The rule-based reasoning engine of the obstetrics knowledge graph is used to verify whether the candidate event conforms to the preset medical common sense, combined with the entity association matching results, and the event confidence adjustment factor is output. The event confidence adjustment factor is multiplied by the event probability corresponding to the candidate event in the event probability vector to obtain the final confidence level, and the corresponding inference result is generated. If the final confidence level exceeds the second threshold, the inference result is a confirmation that the event has occurred. The event type is determined according to the type of the candidate event, and the center time of the feature extraction window corresponding to the candidate event is used as the start timestamp. The estimated end time is determined based on the event duration rules in the obstetrics knowledge graph. The feature extraction window is the audio frame window corresponding to the Hamming window or the acceleration feature extraction window corresponding to the sliding window. The event type, the start timestamp, and the estimated end time are integrated to generate the corresponding event information.

5. The method according to claim 1, characterized in that, The preset Mask R-CNN instance segmentation network includes a region proposal network and a classification branch. The region proposal network includes 16×16 anchor boxes and 32×32 anchor boxes. The frame image is input into a preset Mask R-CNN instance segmentation network to identify sensitive regions of non-newborn subjects; The desensitization strategy is dynamically adjusted based on the event type in the optimized global context object. For the sensitive region, a chaotic sequence is generated based on the UID in the optimized global context object and the current frame timestamp. The sensitive region is then irreversibly masked using a chaotic pixel replacement algorithm, generating a desensitized video frame. The sensitive region category and processing parameters during the desensitization process are recorded, generating desensitization operation metadata, including: The frame image is input into the preset Mask R-CNN instance segmentation network. Multiple candidate region boxes are generated using the 16×16 and 32×32 anchor boxes of the region proposal network. Region pooling is performed on the candidate region boxes to obtain the feature map corresponding to each candidate region box. The feature maps are input into the classification branch of the preset Mask R-CNN instance segmentation network. The classification branch determines the category of the target corresponding to each feature map and predicts the segmentation mask and confidence level of each target. The bounding box, category label, segmentation mask and confidence level of each target are output. Targets with confidence levels exceeding a preset confidence threshold are identified as detection targets. Targets with category labels of medical staff face, department logo or background infant are selected from the detection targets as the sensitive regions. The event type is extracted from the optimized global context object, and the desensitization strategy is dynamically adjusted according to the event type to obtain the target desensitization strategy; the UID in the optimized global context object is concatenated with the current frame timestamp to obtain the concatenation result, and a hash operation is performed on the concatenation result to obtain the initial seed of the chaotic mapping; For each of the aforementioned sensitive regions, the initial seed of the chaotic mapping is iteratively calculated based on the Logistic chaotic mapping formula to generate a chaotic sequence equal to the total number of pixels in the sensitive region, and a permutation index array equal to the total number of pixels in the sensitive region is generated based on the chaotic sequence. The pixels in the sensitive area are expanded into a one-dimensional pixel array in row order, and the pixel positions of the one-dimensional pixel array are rearranged according to the permutation index array to obtain a rearranged one-dimensional pixel array. The rearranged one-dimensional pixel array is filled back into the sensitive region according to the original region coordinates. The filled sensitive region is then merged with the unprocessed region in the frame image to generate the desensitized video frame. Record the category of the sensitive region, the bounding box coordinates, the initial seed of the chaotic mapping, and the generation parameters of the permutation index array to obtain the metadata of the desensitization operation.

6. The method according to claim 5, characterized in that, The mathematical expression of the Logistic chaotic mapping formula is: in, This serves as the initial seed for the chaotic mapping; The first of the UIDs Bit binary value, ; The current frame timestamp is the Bit binary value, ; This is a bitwise XOR operation; For the first The values ​​of the chaotic sequence in the next iteration; and All are piecewise chaos control parameters; and This refers to the category weight adjustment coefficient for sensitive areas; Weights for sensitive region categories; For the first pixel displacement index; This represents the total number of pixels in the sensitive region.

7. A newborn image recording and detection system based on artificial intelligence, characterized in that, The system includes: The identity binding and video acquisition module is used to read the UID and newborn basic information from the wristband when the near-field communication signal of the wristband is detected, acquire the initial video stream in real time through the image acquisition device, obtain the current system timestamp as the binding timestamp, generate a dynamic bloodline tag based on the UID and the binding timestamp, and embed the dynamic bloodline tag into the initial image of each frame of the initial video stream through frequency domain transformation to obtain the bloodline binding video stream. The module also packages the UID, the binding timestamp, the dynamic bloodline tag and the newborn basic information to generate a global context object. An intelligent event triggering module is used to acquire audio streams and accelerometer data streams in real time, extract features from the audio streams and accelerometer data streams to obtain multimodal feature tensors, input the multimodal feature tensors into a preset temporal convolutional network, and output event probability vectors. Based on the event probability vectors, a rule-based reasoning function is performed by querying an obstetrics knowledge graph. When the reasoning result confirms the occurrence of an event, event information is generated, and a collection triggering command is generated based on the event information. The event information is stored in a preset event list, and the preset event list is appended to the global context object to obtain an optimized global context object. The collection triggering command is used to control the start and stop of the acquisition of the initial video stream by the image acquisition device. The event information includes event type, start timestamp, and estimated end time. The privacy protection and desensitization module is used to obtain the timestamp corresponding to each frame of the bloodline-bound video stream as the current frame timestamp, input the frame image into a preset Mask R-CNN instance segmentation network to identify sensitive regions that are not newborn subjects; dynamically adjust the desensitization strategy according to the event type in the optimized global context object; generate a chaotic sequence for the sensitive regions based on the UID in the optimized global context object and the current frame timestamp, irreversibly cover the sensitive regions through a chaotic pixel replacement algorithm to generate desensitized video frames, record the sensitive region categories and processing parameters during the desensitization process, generate desensitization operation metadata, append the desensitization operation metadata to the optimized global context object to obtain the final global context object, and combine each of the desensitized video frames to obtain the desensitized video stream; The file structuring and archiving module is used to read the preset event list from the final global context object, and divide the desensitized video stream into multiple video segments along the time axis according to the start timestamp and estimated end time of each event in the preset event list; generate metadata files for each video segment, extract embedded dynamic lineage tags from the video segments, compare and verify the embedded dynamic lineage tags with the dynamic lineage tags in the final global context object to obtain tag verification results, and if the tag verification results are consistent, package each video segment and the corresponding metadata file into a structured file package, and upload the structured file package to the hospital information system.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images