Intelligent inpatient registration and management system based on face recognition

Abstract

This invention discloses an intelligent inpatient registration and management system based on facial recognition, belonging to the field of medical information technology. The system includes an image acquisition module that captures facial image data of individuals awaiting admission in real time using image acquisition devices deployed at hospital entrances and ward areas; a feature comparison module that compares the facial image data with facial feature templates in an electronic health record database using pixel-by-pixel gradient orientation histograms to initially screen a set of candidates with historical medical records; an occlusion completion module that, when a candidate's facial image is partially occluded, invokes a multi-view pose estimation network to synthesize and complete the texture; and a priority ranking module that generates an initial inpatient registration priority queue based on similarity scores and the current queue length of the department being treated, marking the first-ranked candidate as awaiting registration activation. This invention can improve the accuracy of inpatient registration recognition even with facial occlusion and rationally plan the admission order.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an intelligent inpatient registration and management system based on facial recognition. Background Technology

[0002] Current hospital inpatient registration processes typically rely on manual verification of ID documents and paper application forms. Some systems have introduced facial recognition technology for rapid identity verification. However, in inpatient registration scenarios, individuals awaiting admission often have partially obscured facial images due to wearing masks, changes in hairstyle, or camera angles. Conventional facial recognition methods directly extract features from the entire facial area and perform global matching. The lack of information in the obscured areas severely interferes with the accuracy of feature description, making it impossible to effectively identify patients with historical medical records, necessitating manual verification and slowing down registration efficiency. Existing technologies often handle facial occlusion by discarding obscured samples after detection or treating obscured areas as having zero weight. These methods do not truly restore the lost facial texture information, and when the degree of occlusion is severe, the similarity score drops significantly, causing many candidates who could have been automatically matched to be excluded. In addition, the existing inpatient registration system generally sorts patients based on the order of registration time or a single identity verification result, without linking the reliability of the identity matching to the real-time queuing status of the department they are currently visiting. This can lead to a backlog of patients in urgent need of hospitalization at different departments, and bed resources may not be available in a timely manner.

[0003] This invention aims to address the issues of reduced reliability of hospital registration recognition when faces are obscured and the disconnect between hospital registration priority and the real-time dynamics of the visiting department. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent inpatient registration and management system based on facial recognition, which can still obtain complete recognition similarity even when the face of the person waiting to be admitted is partially obscured, and integrate the similarity with the queue length of the department to generate a dynamic priority queue, thereby improving the response efficiency and resource coordination capabilities of the inpatient registration process.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent inpatient registration and management system based on facial recognition, including an image acquisition module, a feature comparison module, an occlusion completion module, and a priority ranking module. The image acquisition module is used to capture facial image data of people awaiting admission in real time using image acquisition devices deployed at the hospital entrance and ward areas, thereby completing the collection of identity information in a non-intrusive manner and reducing personnel contact and queuing time. The feature comparison module is used to perform pixel-by-pixel gradient orientation histogram comparison of the facial image data with facial feature templates in a pre-stored electronic health record database, initially screening out a set of candidates with historical medical records. Utilizing the high robustness of the gradient orientation histogram to changes in lighting and facial expression, it can quickly identify potential patients and narrow down the scope of identity recognition. The occlusion completion module is used to call a multi-view pose estimation network to synthesize and complete the texture of the occluded area when the facial image data of any person in the candidate set is partially occluded. Based on the completed facial image data, the similarity score with the facial feature template is recalculated, thereby overcoming the loss of facial information caused by masks, glasses, or other occlusions and ensuring that the recognition confidence does not decrease due to sudden occlusion. The priority ranking module is used to generate an initial inpatient registration priority queue based on the similarity score and the current waiting length of each person in the candidate set. The target person at the top of the initial inpatient registration priority queue is marked as pending registration, allowing patients with high identity verification and urgent waiting times to enter the inpatient registration process first, thus optimizing the allocation of medical resources.

[0006] As a preferred embodiment of the present invention, the image acquisition module is specifically used to: control multiple wide-angle cameras installed at the entrance gates and the tops of the ward corridors of the hospital to synchronously trigger shooting actions at fixed time intervals, acquiring original color images containing the head and shoulder areas of the people awaiting admission; perform histogram equalization and Gaussian filtering on the original color images to extract effective facial regions with pixel values ​​within a preset brightness range; and rotate and correct the effective facial regions according to the direction of the line connecting the pupils of both eyes, so that the corrected facial image data is in a uniform horizontal reference posture. Through standardized preprocessing at the same level, comparison errors caused by differences in shooting angle and lighting are eliminated, providing consistent input data for subsequent high-precision feature extraction.

[0007] As a further preferred embodiment of the present invention, the feature comparison module is specifically configured as follows: It reads a standard facial feature template corresponding to each historical patient from the electronic health record database. The standard facial feature template includes gradient orientation histogram descriptors in multiple scale spaces. The captured facial image data is divided into grid cells according to the same scale space. The gradient orientation distribution is statistically analyzed within each grid cell to generate a descriptor to be compared. The chi-square distance between the descriptor to be compared and each standard facial feature template is calculated. Historical patients whose chi-square distance is less than a preset first threshold are included in the candidate set. The combination of multi-scale partitioning and chi-square distance measurement allows the comparison process to have a high tolerance for minor changes in expression and age, while effectively excluding non-target individuals with no historical records, thus improving the reliability and speed of candidate set generation.

[0008] As another preferred technical solution of the present invention, the occlusion completion module is specifically used for: performing facial key point detection on the facial image data of each person in the candidate set, extracting geometric constraint points in the eye area, nose tip, and corner of the mouth, and determining whether there is an occlusion area based on the continuity of the distribution of the geometric constraint points; if there is an occlusion area, then inputting the facial image data into a pre-trained multi-view pose estimation network, which outputs virtual facial texture maps from multiple different viewpoints; selecting the corresponding pixel block in the virtual facial texture map closest to the current acquisition viewpoint and filling it into the occlusion area to form the completed facial image data; and recalculating the normalized cross-correlation between the completed facial image data and the facial feature template to obtain an updated similarity score. This solution predicts the texture of the occluded part through a generative network and uses geometric constraints to ensure a natural transition between the completed area and the surrounding facial structure, so that high similarity matching with the template can still be obtained under occlusion conditions, significantly reducing the probability of missed recognition and false rejection.

[0009] In a preferred embodiment of the present invention, the priority ranking module is specifically used for: obtaining from the hospital's real-time scheduling system the number of patients waiting for their current appointment in the department of each candidate in the candidate set and the average consultation time; calculating the expected waiting time for each candidate; sorting each candidate's similarity score from high to low in a first sort, and sorting the expected waiting time from short to long in a second sort; assigning a first weight value and a second weight value to the first sort and the second sort respectively, and obtaining a comprehensive priority score for each candidate after weighted summation; forming the initial inpatient registration priority queue according to the comprehensive priority scores from high to low, and confirming the candidate at the top of the queue as awaiting registration. This dual-dimensional weighted ranking mechanism considers both the reliability of identity verification and the actual waiting pressure, preventing high-trust patients from being delayed in admission due to long queues, and also preventing those with very short waiting times but questionable identities from occupying bed resources first.

[0010] Preferably, the first weight value and the second weight value are dynamically adjusted according to the overall congestion level of the hospital. The higher the congestion level, the larger the second weight value. When the hospital is in peak hours, the system automatically shifts the weight to the queuing time dimension to prioritize the processing of backlogged patients. When congestion eases, the system strengthens the role of identity verification to ensure the accuracy of retrieving the files of registered patients, thereby dynamically matching different operational scenarios.

[0011] In a further improved technical solution of the present invention, the system further includes a bed pre-allocation module for allocating hospital beds to the target personnel in the pending registration state. Specifically, this involves: reading the current list of available beds from the hospital bed management system, the list of available beds including the floor location, room orientation, and nursing level attributes of each bed; obtaining the inpatient medical order information of the target personnel, the inpatient medical order information including the required nursing level and isolation requirements; matching the required nursing level and isolation requirements with the nursing level attributes and room orientation of each bed in the list of available beds to filter out a set of candidate beds that meet the matching conditions; and selecting the bed whose floor location is closest to the department of the target personnel from the set of candidate beds as the pre-allocated bed. By automatically matching medical orders with bed attributes and selecting the nearest floor, the distance for patients to move between departments can be reduced, while ensuring that nursing and isolation requirements are met in a timely manner, avoiding omissions that may occur with manual allocation.

[0012] In another preferred embodiment of the present invention, the system further includes a motion trajectory tracking module, used to monitor the facial movement trajectory of the target person in real time after the target person is marked as pending registration and bed pre-allocation is completed, and to determine whether the target person has actually arrived at the floor where the pre-allocated bed is located. Specifically, the module tracks a sequence of facial image frames of the target person continuously acquired by at least three fixed-point cameras deployed on the floor where the pre-allocated bed is located, calculates the pupil center coordinate offset between adjacent image frames, and, based on the pupil center coordinate offset and the calibration parameters of the fixed-point cameras, reverse-calculates the real-time spatial coordinates of the target person on the floor plan. The real-time spatial coordinates are compared with the bed coordinates corresponding to the pre-allocated bed using Euclidean distance; when the Euclidean distance is less than a preset arrival determination distance, it is confirmed that the target person has actually arrived at the pre-allocated bed. Multi-view vision-based facial trajectory tracking eliminates the need for additional wearable devices, enabling seamless verification of whether a patient has actually entered the ward area. Once arrived, subsequent registration is automatically triggered, reducing the workload of nursing staff in manually confirming bed arrival and preventing bed vacancy.

[0013] In another preferred embodiment of the present invention, the system further includes an automatic registration information filling module, used to initiate an automatic inpatient registration information filling process after confirming that the target person has actually arrived at the pre-assigned bed. Specifically, this involves: retrieving the target person's historical basic personal information from the electronic health record database, including name, gender, date of birth, allergic drugs, and past medical history; obtaining the current system time as the registration time; combining the bed number of the pre-assigned bed, the registration time, and the historical basic personal information into an initial registration electronic form; and calling the inpatient registration information verification rule base, which includes rules for the integrity of required fields and the legality of data formats, to perform item-by-item verification on the initial registration electronic form, generating a verified formal inpatient registration record. This achieves full automation from bed arrival to medical record information filling, eliminating errors in key information such as name and allergy history caused by secondary transcription, and significantly improving the accuracy and efficiency of admission records.

[0014] The system provided by this invention preferably includes an encrypted storage and credential generation module, used to encrypt and store the formal hospitalization registration record and generate an electronic hospitalization credential bound to the facial feature template of the target person. Specifically, this involves: performing a hash algorithm on the person's identity field in the formal hospitalization registration record to obtain a fixed-length digest value as an encrypted index key; storing the formal hospitalization registration record in a blockchain node according to the encrypted index key, and associating the encrypted index key with the facial feature template of the target person; generating an electronic hospitalization credential in the form of a two-dimensional graphic code based on the encrypted index key, and pushing the electronic hospitalization credential to the target person's mobile terminal. By leveraging the tamper-proof characteristics of blockchain to protect the security of hospitalization records, the facial feature-bound credential allows for quick identity verification through facial recognition in subsequent procedures such as medication collection and examinations, eliminating the hassle of repeatedly presenting physical documents, while also ensuring that privacy data is not accessed without authorization.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Facial landmark detection is performed on facial image data from the candidate set. Occlusion areas are determined based on the continuity of geometric constraint point distribution in the periorbital, nasal tip, and mouth corner regions. When occlusion exists, the facial image is input into a pre-trained multi-view pose estimation network, which outputs virtual facial texture maps from multiple different viewpoints. The corresponding pixel block in the texture map closest to the current acquisition viewpoint is selected to fill the occluded area, forming a complete facial image. This scheme uses a multi-view pose estimation network to synthesize virtual textures from different observation angles, rather than simple interpolation or neighborhood filling. It can directly generate textures for missing regions consistent with the current pose, ensuring that the synthesized information of the occluded area maintains viewpoint continuity with the surrounding real facial features. When performing normalized cross-correlation calculations, the completed facial image can provide a descriptor distribution highly similar to that under unoccluded conditions to the facial feature template. This effectively avoids historical medical record matching failures caused by common occlusions such as masks and sunglasses, and even with a large occlusion area, stable similarity scores can be obtained. The system retrieves the number of patients waiting for their current appointment and the average consultation time for each candidate from the hospital's real-time dispatch system, calculating the estimated waiting time. The similarity scores are then ranked from highest to lowest, followed by the estimated waiting time from shortest to longest. Weights are assigned to both rankings, and a weighted sum is calculated to obtain a comprehensive priority score. This score forms the initial priority queue for inpatient registration. This method incorporates both the confidence level of facial recognition matching and the actual queue length in each department into the ranking criteria. This ensures that individuals with more reliable matching and shorter waiting times are prioritized for inpatient registration, rather than relying solely on registration time or a single identity score. This avoids situations where uncertain match identities are prioritized, or highly matched individuals are registered early but cannot be admitted promptly due to departmental congestion. The inpatient registration order dynamically aligns with the actual departmental operational rhythm, reducing the ineffective occupancy of beds after registration. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a schematic diagram of the structure of an intelligent inpatient registration and management system based on facial recognition; Figure 2 This is a flowchart of the facial recognition comparison process for hospital admissions. Figure 3 This is a flowchart of face recognition processing based on partial occlusion repair and multi-view pose estimation; Figure 4This is a flowchart of the inpatient bed pre-allocation and arrival confirmation process; Figure 5 This is a flowchart of the electronic hospitalization certificate generation method. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides an intelligent inpatient registration and management system based on facial recognition, including an image acquisition module, a feature comparison module, an occlusion completion module, and a priority ranking module. The image acquisition module captures facial image data of individuals awaiting admission in real time using image acquisition devices deployed at the hospital entrance and ward areas. The feature comparison module performs pixel-by-pixel gradient orientation histogram comparison between the facial image data and facial feature templates in a pre-stored electronic health record database to initially filter out a set of candidates with historical medical records. When the facial image data of any individual in the candidate set is partially occluded, the occlusion completion module calls a multi-view pose estimation network to synthesize and complete the texture of the occluded area, and recalculates the similarity score with the facial feature template based on the completed facial image data. The priority ranking module generates an initial inpatient registration priority queue based on the similarity score and the current queuing length of each department in the candidate set, and marks the first-ranked candidate in the initial inpatient registration priority queue as awaiting registration activation. Example 1:

[0020] In specific implementation, please refer to Figure 2 The system controls multiple wide-angle cameras installed at the hospital's entrance turnstiles and along the ceilings of ward corridors. These cameras synchronously trigger shooting at fixed time intervals to acquire raw color images including the head and shoulder areas of individuals awaiting admission. The cameras are positioned to completely cover the turnstile and ward corridor areas. The preset time interval of 0.5 seconds ensures that at least three frames of facial views are captured during passage. Synchronous shooting is triggered by pulse signals emitted from a central control unit connected to each wide-angle camera. The rising edge of the pulse signal simultaneously activates the electronic shutters of all wide-angle cameras.

[0021] After acquiring the original color image, histogram equalization and Gaussian filtering denoising are performed. Histogram equalization maps the luminance channels of the original color image using a cumulative distribution function, resulting in an approximately uniform distribution of the luminance histogram within the 0-255 grayscale range, thus enhancing local contrast in shadow or highlight areas. Gaussian filtering denoising uses a 5×5 pixel Gaussian convolution kernel with a standard deviation parameter set to 1.0. This suppresses random noise generated by the image sensor in low-light environments by weighted averaging of the neighboring pixel values ​​of each pixel in the original color image. After histogram equalization and Gaussian filtering denoising, effective facial regions with pixel values ​​within a preset luminance range are extracted. The preset luminance range is set to 60-240 luminance values. Pixels below 60 are considered overly dark backgrounds or occlusions, and pixels above 240 are considered overexposed areas. Only connected components with luminance values ​​within this range are retained, and the connected component with the largest area is selected as the effective facial region.

[0022] After extracting the effective facial region, the region is rotated and corrected according to the direction of the line connecting the pupils of both eyes. First, a cascaded regression facial keypoint detection algorithm is used to locate the center coordinates of the left and right pupils within the effective facial region. Based on these coordinates, the angle between the direction of the line connecting the pupils and the horizontal direction is calculated. Using the midpoint of these coordinates as the rotation center, an affine transformation matrix is ​​used to rotate the effective facial region in the opposite direction, with the rotation angle equal to the aforementioned angle. This ensures that the corrected line connecting the pupils is horizontal, thus bringing the corrected facial image data to a uniform horizontal reference pose.

[0023] The standard facial feature template for each historical patient is retrieved from the electronic health record database. This template includes gradient orientation histogram descriptors across multiple scales. The standard facial feature templates for historical patients pre-stored in the electronic health record database are generated as follows: During the patient's first visit, a standard frontal facial image is captured using a wide-angle camera of the same specifications. This image is then divided into grid cells across multiple scales. Each scale is used to construct an image pyramid hierarchy using Gaussian kernels with different blur coefficients (e.g., using kernels with blur coefficients σ of 1.0, 1.5, and 2.0 to generate three levels). Each level of the image is then divided into 16×16 pixel grid cells, with an 8-pixel overlap between grid cells. Within each grid cell, the gradient direction and magnitude of each pixel are calculated. The gradient direction value is quantized into nine directional intervals. The gradient direction distribution is statistically analyzed, and a gradient orientation histogram descriptor is generated. The gradient orientation histogram descriptors of all grid cells are concatenated in spatial order to form the standard facial feature template.

[0024] The captured facial image data is spatially divided into grid cells of the same scale. Specifically, the same image pyramid hierarchy and grid cell division method as the standard facial feature template are used, i.e., three levels are constructed using Gaussian kernels with blur coefficients σ of 1.0, 1.5, and 2.0, and each level is divided into 16×16 pixel grid cells, with an 8-pixel overlap between grid cells. The gradient direction distribution is statistically analyzed within each grid cell to generate a comparison descriptor. Within each grid cell, the Sobel operator is used to calculate the horizontal and vertical gradient components of each pixel. The gradient direction value and gradient magnitude are calculated from the horizontal and vertical gradient components. The gradient direction value is quantized into 9 directional intervals. The gradient magnitude is used as the weight to calculate the weighted sum of the gradient magnitudes within each directional interval, resulting in the gradient direction histogram descriptor for that grid cell. The gradient direction histogram descriptors of all grid cells are concatenated in the same spatial order to form a comparison descriptor with the same dimensions as the standard facial feature template.

[0025] Calculate the chi-square distance between the descriptor to be compared and each standard face feature template. The chi-square distance is calculated as follows:

[0026] in, This represents the chi-square distance value between the descriptor to be compared and the current standard face feature template. This represents the total number of dimensions of the descriptors to be compared. The index number describes the sub-dimension, and its value ranges from 1 to... Integers between [a certain number] Indicates the descriptor to be compared in the th... Statistical values ​​of gradient directions in each dimension The standard face feature template is shown in the first position. Gradient direction statistics in each dimension. When and When the chi-square distance is zero, the contribution of this dimension to the chi-square distance is set to zero. Historical patients whose chi-square distance is less than a preset first threshold are included in the candidate set. The preset first threshold is determined by statistically analyzing the distribution of chi-square distances of multiple facial image comparisons of the same person in the historical patient sample database. Specifically, it is set to the upper quartile of this distribution to ensure that true matching pairs are fully preserved while filtering out obviously irrelevant historical patients. Example 2:

[0027] In specific implementation, please refer to Figure 3Facial landmark detection was performed on the facial image data of each candidate in the candidate set. Facial landmark detection was implemented using a shape prediction model based on cascaded regression trees. This shape prediction model includes an initial shape estimator and a five-level cascaded regressor, each level consisting of twenty gradient boosting regression trees, with a maximum depth of four layers for each tree. After converting the facial image data to grayscale, it was input into the shape prediction model, which output the coordinates of sixty-eight facial landmarks, including six landmarks for the left eye, six landmarks for the right eye, four landmarks for the bridge of the nose, one landmark for the tip of the nose, twelve landmarks for the outer contours of the upper and lower lips, and two landmarks for the left and right corners of the mouth. From these sixty-eight facial landmark coordinates, twelve landmarks in the periorbital region were extracted as periorbital geometric constraint points, the landmark at the tip of the nose was extracted as a nasal tip geometric constraint point, and the landmarks at the left and right corners of the mouth and the center landmarks of the upper and lower lips were extracted as geometric constraint points for the corners of the mouth.

[0028] When determining the existence of occlusion areas based on the continuity of the distribution of geometric constraint points, the Euclidean distance between adjacent key points in the geometric constraint points around the left eye and the right eye is calculated. If either Euclidean distance exceeds a preset periorbital distance threshold, the periorbital area is considered occluded. The preset periorbital distance threshold is determined based on historical data of the average distance of key points around the eyes of the same person under unoccluded conditions, and is set to 1.2 times the maximum value of this distance in the unoccluded sample. For the geometric constraint points at the tip of the nose, the distance between the geometric constraint point at the tip of the nose and the highest key point among the four key points on the bridge of the nose is calculated. If this distance exceeds a preset bridge of the nose distance threshold, the nasal area is considered occluded. For the geometric constraint points at the corners of the mouth, the horizontal distance between the key points at the left and right corners of the mouth is calculated. If this horizontal distance is less than a preset corner of the mouth distance threshold, the mouth area is considered occluded. The preset corner of the mouth distance threshold is set to 0.8 times the minimum value of this horizontal distance under unoccluded conditions. When at least one of the eye area, nose area, or mouth area is determined to be occluded, a local occlusion area is identified in the facial image data, and the areas determined to be occluded are merged into an occlusion area.

[0029] If occlusion exists, the facial image data is input into a pre-trained multi-view pose estimation network. The multi-view pose estimation network adopts an encoder-decoder architecture. The encoder consists of six convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The kernel size is 3×3. The stride is set to 2 in the first, third, and fifth convolutional blocks to achieve downsampling, and the stride of the remaining convolutional blocks is 1. The number of channels is 64, 128, 256, 256, 256, and 256 respectively, and the final output is a 16×16×256 feature code. The multi-view pose estimation network comprises five parallel decoder branches, each corresponding to a preset output viewpoint. These five preset output views are: a frontal viewpoint with a horizontal yaw angle of 0° and a vertical pitch angle of 0°; a left-side viewpoint with a horizontal yaw angle of +30° and a vertical pitch angle of 0°; a right-side viewpoint with a horizontal yaw angle of -30° and a vertical pitch angle of 0°; a top-down viewpoint with a horizontal yaw angle of 0° and a vertical pitch angle of +15°; and a top-down viewpoint with a horizontal yaw angle of 0° and a vertical pitch angle of -15°. Each decoder branch consists of six upsampled convolutional blocks, each containing a transposed convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The transposed convolutional layer has a 3×3 kernel size and a stride consistent with the downsampling stride of the corresponding layer in the encoder, used to progressively restore spatial resolution. Finally, each decoder branch outputs a three-channel virtual facial texture map with the same resolution as the input facial image data.

[0030] The training process of the multi-view pose estimation network is as follows: A training sample set is constructed. Each training sample in the set consists of a face image with random partial occlusion, a corresponding unoccluded real face image, and five preset viewpoint texture maps corresponding to that real face image. The face image with random partial occlusion is obtained by randomly adding black rectangular blocks, circular blocks, or irregular polygonal blocks to a multi-view face image generated from a publicly available 3D face model dataset. The face image with random partial occlusion is used as the network input, and the five preset viewpoint texture maps are used as the network output targets. A multi-task loss function is used for training. The multi-task loss function is obtained by weighted summation of pixel-level L1 loss and perceptual loss, where the weight of pixel-level L1 loss is set to 1.0, and the weight of perceptual loss is set to 0.01. The perceptual loss is calculated using the L2 distance between the feature maps output from layers conv3_3 and conv4_3 of a VGG-19 network pre-trained on the ImageNet dataset. The optimizer used is the Adam optimizer, with an initial learning rate of 0.0002. After every ten training iterations, the learning rate is reduced to 0.8 times its original value. The batch size is set to 8. Training continues until the pixel-level L1 loss no longer decreases after five consecutive iterations on the validation set.

[0031] When selecting the corresponding pixel block in the virtual facial texture map closest to the current acquisition viewpoint, the current acquisition viewpoint is first estimated using a head pose estimation algorithm on the facial image data with occluded areas. This algorithm solves the perspective problem of n points between multiple key points on the facial image data and their corresponding points in a general 3D face model, obtaining the horizontal deflection angle and vertical pitch angle of the current acquisition viewpoint. The first absolute value of the difference between the horizontal deflection angle of the current acquisition viewpoint and the horizontal deflection angles of five preset output viewpoints, and the second absolute value of the difference between the vertical pitch angle of the current acquisition viewpoint and the vertical pitch angles of the five preset output viewpoints are calculated. The output viewpoint with the smallest sum of the first and second absolute values ​​is determined as the closest viewpoint, and the virtual facial texture map corresponding to this closest viewpoint is selected as the filling source map. In the filling source map, pixel blocks within the same coordinate range are extracted based on the geometric constraint point coordinate range corresponding to the occluded area. These pixel blocks are then applied pixel by pixel to the occluded area of ​​the facial image data. During the application, a linear feathering fusion of α pixel width is performed on the edges of the pixel blocks. α is set to 2 to smoothly stitch the boundaries, forming the complete facial image data after completion.

[0032] After generating the complete facial image data, the complete facial image data is re-correlated with the facial feature template using normalized cross-correlation. The normalized cross-correlation calculation formula is as follows:

[0033] in, This indicates the facial feature template in pixel coordinates. Pixel value at that location, This represents the mean value of all pixels in the facial feature template. This indicates the pixel coordinates of the completed facial image data after completion. Pixel value at that location, This represents the mean of all pixel values ​​in the completed facial image data after padding. This represents the summation of all pixel coordinates within the overlapping area of ​​the completed facial image data and the facial feature template. Before calculation, the completed facial image data is scaled to the same size and horizontal reference pose as the facial feature template. The calculated result... The updated similarity score. The value range is from -1 to 1. The closer the value is to 1, the higher the similarity between the completed facial image data and the facial feature template. Example 3:

[0034] In practice, the system retrieves the waiting number and average consultation time for each candidate from the hospital's real-time dispatch system, corresponding to their currently registered department. The hospital's real-time dispatch system provides real-time queue data for each department via an HTTP interface. The interface returns a data structure containing a unique department identifier, the current waiting number, and the average consultation time over the past 30 minutes, expressed in minutes and rounded to one decimal place. Each candidate is associated with their registered information through their personal identification recorded in the electronic health record database, thus obtaining a unique identifier for their registered department. Based on this unique identifier, the system then queries the hospital's real-time dispatch system interface for the corresponding waiting number and average consultation time.

[0035] The formula used to calculate the estimated waiting time for each person is:

[0036] in, This represents the estimated waiting time for the currently selected candidate in the candidate set, in minutes. This indicates the number of people currently waiting for treatment in the department that the person is registered with in the hospital's real-time dispatch system. This indicates the average consultation time (in minutes) for the department the person registered with in the hospital's real-time dispatch system over the past 30 minutes. If the average consultation time interface does not return a valid value, then... The default value is assigned based on the historical average consultation time of the department. The historical average consultation time is obtained by taking the arithmetic average of the consultation records of the department in the same time period over the past thirty days.

[0037] The similarity scores of each person are sorted from highest to lowest, and the sorting results are represented by natural number indices. The first sorting number is the first number of the person with the highest similarity score. The first ranking number of the person with the second highest similarity score, which is 1. The initial ranking number is 2, and subsequent numbers increase sequentially. If there are individuals with the same similarity score, they are assigned the same initial ranking number, and subsequent numbers are deferred. The estimated waiting time for each individual is then sorted in ascending order, with the sorting results represented by natural numbers. The individual with the shortest estimated waiting time is assigned the second ranking number. The second ranking number is 1, indicating the person with the second shortest expected waiting time. The first number is 2, and subsequent numbers increase sequentially. If there are cases where the expected waiting time is the same, the same second sorting number is assigned to people with the same expected waiting time, and subsequent numbers are deferred.

[0038] The first and second rankings are assigned a first weight value and a second weight value, respectively. The overall congestion level of the hospital is calculated by dividing the total number of patients waiting in each department by the total number of clinics in the hospital. This ratio is denoted as the congestion proportion. A congestion proportion less than 0.5 is considered mild congestion, between 0.5 and 1.0 is considered moderate congestion, and greater than 1.0 is considered severe congestion. In the mild congestion level, the first weight value is set to 0.7, and the second weight value is set to 0.3. In the moderate congestion level, both the first and second weight values ​​are set to 0.5. In the severe congestion level, the first weight value is set to 0.3, and the second weight value is set to 0.7. The sum of the first and second weight values ​​is always 1.0.

[0039] The weighted summation yields the overall priority score for each individual. The formula for calculating the weighted summation is as follows:

[0040] in, This indicates the overall priority score of the current calculator. This represents the first weight value. This indicates the first sorting number of the current calculator. This represents the second weight value. This represents the second sorting number of the current calculator. Since both the first and second sorting numbers are considered better the smaller they are, a comprehensive priority score is applied. The smaller the value, the higher the priority of the person.

[0041] An initial inpatient registration priority queue is formed by arranging patients in descending order of their comprehensive priority scores. The comprehensive priority scores are used during the ranking process. The person with the lowest value is placed at the front of the queue, based on a comprehensive priority score. The person with the second lowest score is placed second in the queue, and so on. If a comprehensive priority score exists... In cases where similarity scores are the same, individuals with higher similarity scores are ranked higher. The target individual at the top of the initial inpatient registration priority queue is identified as being in a pending registration status, and a notification of pending registration status and inpatient registration preparation instructions are pushed to the target individual's mobile device. At the same time, the target individual's identity and queue ranking result are written into the registration status cache of the hospital management system. Example 4:

[0042] In specific implementation, please refer to Figure 4After marking the top-ranked candidate in the initial inpatient registration priority queue as pending registration, inpatient beds are allocated to these candidates. The system reads the current list of available beds from the hospital bed management system, which maintains a relational database table containing fields for bed number, floor location, room orientation, and nursing level. The floor location field stores a combined code of the bed's floor number and room number. The floor number is represented by two digits from basement level to floor 15, and the room number by three digits. The room orientation field takes one of the following directions: east, south, west, north, southeast, northeast, southwest, or northwest. The nursing level field takes one of three values: Level 1, Level 2, or Level 3, indicating the highest nursing level the bed can accommodate. The current list of available beds is obtained by executing a query in the hospital bed management system where the bed occupancy status field equals "available," returning the bed number, floor location, room orientation, and nursing level of all beds that meet the condition.

[0043] When obtaining the inpatient medical order information of a target individual, the inpatient medical order record filled out by the attending physician with prescription authority during this visit is retrieved from the hospital's electronic medical record system. The inpatient medical order information includes two fields: required nursing level and isolation requirement. The required nursing level field takes one of three values: Level 1, Level 2, or Level 3 nursing care. The isolation requirement field takes one of two values: required isolation or not required isolation. The required nursing level is determined by the attending physician based on the severity of the target individual's condition and their ability to perform daily living activities, as assessed in the inpatient medical order. The isolation requirement is marked as required isolation based on the target individual's infectious disease screening results, specifically a positive result in either multidrug-resistant bacteria testing or respiratory infectious disease testing.

[0044] When matching the required nursing care level and isolation requirements with the nursing care level attribute and room orientation of each bed in the list of available beds, the nursing care level matching condition is that the bed's nursing care level attribute must be greater than or equal to the required nursing care level. The order of nursing care levels is: Level 1 nursing care is higher than Level 2 nursing care, and Level 2 nursing care is higher than Level 3 nursing care. The isolation requirement matching condition is that when the isolation requirement is required, the room orientation must be south, southeast, or southwest, and the bed must be located on the top or second-to-top floor of the inpatient building to meet the requirements of natural ventilation and relative independence. The top floor refers to the highest floor, and the second-to-top floor refers to a floor only lower than the highest floor. When the isolation requirement is not required, there is no restriction on the room orientation. Available beds that simultaneously meet the nursing care level matching condition and the isolation requirement matching condition are filtered out to form a candidate bed set.

[0045] The bed closest to the target person's department in the candidate bed set is selected as the pre-assigned bed. The department is located within the hospital's outpatient building. The geometric center coordinates of each department and the room geometric center coordinates of each bed in the vacant bed list are pre-marked using a hospital building map. The straight-line distance between the geometric center coordinates of the department and each candidate bed (in meters) is calculated. The candidate bed with the smallest straight-line distance is selected as the pre-assigned bed. If multiple candidate beds have the same straight-line distance, the one on the lower floor is prioritized. After determining the pre-assigned bed, the occupancy status of the bed in the hospital bed management system is updated to "pre-assigned," and the target person's identity is associated with the bed number and written into the pre-assignment record table.

[0046] After bed pre-allocation is completed, the facial image frame sequence of the target person is continuously captured by at least three fixed-point camera devices deployed on the floor where the pre-allocated bed is located. The three fixed-point camera devices are respectively installed above the elevator exit ceiling, at the end of the floor corridor centerline, and on the wall opposite the floor nurse station. Each fixed-point camera uses a 1920×1080 pixel color camera with a frame rate of 15 frames per second. After the target person is marked as pending registration, the facial feature template of the target person and the floor information of the pre-allocated bed are read from the pre-allocation record table. The facial feature template of the target person is then sent to the edge computing nodes corresponding to the three fixed-point camera devices. Lightweight face detection and tracking models are deployed on the edge computing nodes.

[0047] Edge computing nodes detect the target person's face from received consecutive image frames. After detecting the target person's face, they generate a sequence of image frames containing the target person's face. When calculating the pupil center coordinate offset between adjacent image frames, the same facial key point detection algorithm as in Example 2 is used to extract the left and right pupil center coordinates for each frame. The average of the left and right pupil center coordinates is taken as the pupil center coordinates for that frame. Two adjacent frames in the image frame sequence are defined as the [number missing]. Frame and the Frame, number The pupil center coordinates of the frame are as follows , No. The pupil center coordinates of the frame are as follows The pupil center coordinate offset is composed of the horizontal offset component. and vertical offset components composition.

[0048] Based on the pupil center coordinate offset and the calibration parameters of the fixed-point camera devices, the real-time spatial coordinates of the target person on the floor plan are deduced. The three fixed-point camera devices had their internal and external parameters calibrated during installation; the internal parameters included focal length. Principal point coordinates The external parameters include the rotation matrix and translation vector, and the calibration parameters are obtained using Zhang Zhengyou's calibration method. Using the first... The pupil center coordinates are simultaneously acquired by at least two fixed-point cameras. The three-dimensional spatial coordinates of the target person in the floor coordinate system are calculated using binocular geometric triangulation. These three-dimensional spatial coordinates are then projected onto a two-dimensional floor plan map to obtain the [further details needed]. Real-time spatial coordinates corresponding to the frame During projection, the height coordinates are discarded. The real-time spatial coordinates are compared with the bed coordinates corresponding to the pre-assigned beds using Euclidean distance. The formula for calculating Euclidean distance is:

[0049] in, This represents the Euclidean distance between the real-time spatial coordinates and the coordinates of the bed corresponding to the pre-assigned bed. This represents the real-time spatial x-coordinate of the target personnel on the floor plan map. This represents the real-time spatial coordinates of the target personnel on the floor plan. This indicates the horizontal coordinate of the pre-assigned bed on the floor plan. This indicates the vertical coordinate of the pre-assigned bed on the floor plan. The preset reachability distance is set based on the typical distance from the ward entrance to the head of the bed, using an empirical value of 2.5 meters. (This is based on the Euclidean distance.) If the distance is less than 2.5 meters, confirm that the target person has actually arrived at the pre-assigned bed. Example 5:

[0050] See Figure 5After confirming that the target person has actually arrived at the pre-assigned bed, the automatic inpatient registration information filling process is initiated. The target person's historical basic personal information is retrieved from the electronic health record database. This historical basic personal information includes name, gender, date of birth, allergic drugs, and past medical history. The electronic health record database is deployed on the hospital's intranet core database server, using a relational database for storage. Historical basic personal information is stored in a personal basic information table. This table uses the person's identity identifier as the primary key. The name field is a Unicode string of up to 20 characters; the gender field is a single character with the value "male" or "female"; the date of birth field is a date type in YYYY-MM-DD format; the allergic drugs field is a comma-separated list of drug names of up to 200 characters; and the past medical history field is a comma-separated list of disease codes of up to 500 characters. Using the target person's identity identifier, a parameterized SELECT query is performed in the personal basic information table of the electronic health record database, returning the corresponding values ​​for the name, gender, date of birth, allergic drugs, and past medical history fields.

[0051] The current system time is obtained as the registration time. The system time is synchronized from the hospital's unified time-synchronized NTP server, with the time zone set to UTC+8, time precision in seconds, and record format YYYY-MM-DDHH:MM:SS. The bed number of the pre-assigned bed, the registration time, and historical personal basic information are combined into an initial electronic registration form. The bed number of the pre-assigned bed is read from the pre-assignment record table, which has already written the correspondence between bed numbers and target personnel during the bed allocation stage. A hierarchical Extensible Markup Language (XML) document is constructed during the combination process. The root element is inpatient registration information, containing four child elements: the bed information child element contains a bed number child element; the registration time child element has a value of the registration time string; the personal information child element contains five atomic elements: name, gender, date of birth, allergic drugs, and past medical history; and the historical archive child element is used to indicate that the record originates from the electronic health record database.

[0052] The system invokes a rule base for verifying inpatient registration information. This rule base includes rules for the integrity of required fields and the validity of data formats. The rule base is implemented using a rule engine, which predefines several verification rules and executes them sequentially according to their priority. The rules for the integrity of required fields are as follows: the name element cannot be an empty string; the gender element must be either male or female; the date of birth element must be a valid date with the corresponding age between 0 and 150 years old; and the bed number element must match one of the available or pre-assigned bed numbers in the hospital's bed management system. The rules for the validity of data formats are as follows: the name element is only allowed to contain Chinese characters and uppercase and lowercase English letters; the allergy drug element, if not empty, must consist of standard generic drug names separated by semicolons, and each generic name must have a corresponding entry in the hospital's drug dictionary; and the past medical history element, if not empty, must consist of comma-separated International Classification of Diseases 10th Revision (ICD-10) codes, and each code must have a corresponding entry in the disease classification coding table. The rules engine checks each element of the initial registration electronic form against the rules. If any check fails, an error report is generated and sent back to the inpatient registration management terminal for correction. If all checks pass, a valid inpatient registration record is generated. The valid inpatient registration record adds a check pass flag element and a check timestamp element to the initial registration electronic form.

[0053] After generating the official hospitalization registration record, the record is encrypted and stored, and an electronic hospitalization certificate is generated that is bound to the target individual's facial feature template. A hash algorithm is used to perform a digest operation on the person's identification field in the official hospitalization registration record, resulting in a fixed-length digest value used as the encryption index key. The person's identification field in the official hospitalization registration record is taken from the target individual's identification in the electronic health record database and is an alphanumeric string with a maximum length of eighteen characters. The hash algorithm used is a 256-bit version of the Secure Hash Algorithm, and the digest operation calculation formula is as follows:

[0054] in, This represents the encrypted index key obtained after hashing the personnel identification field; it is a fixed-length string consisting of 64 hexadecimal characters. This represents the 256-bit version of the secure hash algorithm, which internally undergoes message padding, message expansion, and sixty-four rounds of compression iterations. This represents the raw string value of the person's identification field in the official hospitalization registration record. Before calculation... Convert a string to a UTF-8 encoded byte sequence.

[0055] Formal hospital admission records are stored in the blockchain nodes according to an encrypted index key. The blockchain adopts a consortium blockchain architecture based on practical Byzantine fault tolerance. The consortium blockchain network consists of a hospital core information node, a medical affairs management node, and a data security audit node, and all nodes communicate via TLS encrypted communication. When storing formal hospital admission records, the entire formal hospital admission record is used as the storage payload. It is first encrypted using the symmetric encryption algorithm Advanced Encryption Standard 256-bit in Galois / counter mode. The encryption key is derived by the hospital key management server based on the master key and the encrypted index key through a key derivation function. The encrypted ciphertext and the encrypted index key are encapsulated together as a data storage transaction. The transaction is submitted to the consortium blockchain network by the hospital core information node. After the consensus nodes in the consortium blockchain network execute the three-phase consensus protocol, they write the data storage transaction into the new block of the blockchain ledger maintained by each node, achieving tamper-proof and traceable encrypted storage.

[0056] The encrypted index key is associated with the facial feature template of the target individual. The facial feature template is stored in a facial biometric database, a non-relational key-value store where the key is the feature template identifier corresponding to the facial feature template, and the value is the feature description data. During the association process, the facial feature template record corresponding to the target individual is searched in the facial biometric database. An extended attribute field is added to this record, named "Hospitalization Certificate Index," with the encrypted index key as its value. Simultaneously, in the blockchain node, the identifier of the facial feature template is stored as a reverse index using the encrypted index key, enabling quick retrieval of the corresponding facial feature template identifier via the encrypted index key.

[0057] An electronic hospital admission certificate in the form of a two-dimensional graphic code is generated based on the encrypted index key. The Fast Response Matrix Code standard is used during generation, with the encrypted index key string as the encoded content. The error correction level is set to H, which recovers approximately 30% of codeword errors. A version number is selected to ensure that the number of modules in the generated two-dimensional graphic code can accommodate the entire encoded content with a margin. An open-source QR code encoding library is used to perform data analysis, data encoding, error correction encoding, module arrangement, and pattern generation steps, outputting a PNG format two-dimensional graphic code image as the electronic hospital admission certificate.

[0058] The electronic hospitalization certificate is pushed to the target individual's mobile device. The target individual's mobile device identifier is retrieved from the personal contact information table in the electronic health record database. This table contains the individual's identity identifier and a mobile device push token field, which is the registration token for the hospital's official mobile application. Through the hospital's message push middleware, the push service interface corresponding to the mobile operating system of the mobile device push token is called to send a JSON message containing a download link for the electronic hospitalization certificate image and the certificate number to the target individual's mobile device. Upon receiving the message, the mobile device displays a notification indicating that hospitalization registration is complete, and a zoomable 2D graphic code electronic hospitalization certificate is displayed within the hospital's official mobile application.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An intelligent inpatient registration and management system based on facial recognition, characterized in that, include: The image acquisition module is used to capture facial image data of people waiting to be admitted to the hospital in real time through image acquisition devices deployed at the hospital entrance and ward areas; The feature comparison module is used to perform pixel-by-pixel gradient orientation histogram comparison between the facial image data and the facial feature templates in the pre-stored electronic health record database, so as to initially screen out a set of candidates with historical medical records. The occlusion completion module is used to call a multi-view pose estimation network to perform texture synthesis and completion of the occluded area when the facial image data of any person in the candidate set is partially occluded, and to recalculate the similarity score with the face feature template based on the completed facial image data. The priority ranking module is used to generate an initial inpatient registration priority queue based on the similarity score and the current queuing length of each person in the candidate set, and to mark the target person who ranks first in the initial inpatient registration priority queue as a person to be activated for registration.

2. The intelligent inpatient registration and management method based on facial recognition according to claim 1, characterized in that, By deploying image acquisition devices at hospital entrances and ward areas, facial image data of people awaiting admission is captured in real time. Specifically: Control multiple wide-angle cameras installed at the entrance gates and the top of the ward corridors of the hospital to synchronously trigger shooting actions at fixed time intervals to obtain original color images containing the head and shoulder areas of people waiting to be admitted; The original color image is subjected to histogram equalization and Gaussian filtering for noise reduction to extract the effective facial regions whose pixel values ​​are within a preset brightness range. The effective facial region is rotated and corrected according to the direction of the line connecting the pupils of both eyes, so that the corrected facial image data is in a uniform horizontal reference posture.

3. The intelligent inpatient registration and management method based on facial recognition according to claim 2, characterized in that, The facial image data is compared pixel-by-pixel gradient orientation histogram with facial feature templates in a pre-stored electronic health record database to initially filter out a set of candidates with historical medical records. Specifically: Read the standard facial feature template corresponding to each historical patient from the electronic health record database. The standard facial feature template includes gradient orientation histogram descriptors in multiple scale spaces. The captured facial image data is divided into grid cells of the same scale, and the gradient direction distribution is statistically analyzed within each grid cell to generate a descriptor to be compared. Calculate the chi-square distance between the descriptor to be compared and each of the standard facial feature templates, and include historical patients whose chi-square distance is less than a preset first threshold into the candidate set.

4. The intelligent inpatient registration and management method based on facial recognition according to claim 3, characterized in that, When the facial image data of any person in the candidate set is partially occluded, a multi-view pose estimation network is invoked to synthesize and complete the texture of the occluded area. Based on the completed facial image data, the similarity score with the facial feature template is recalculated. Specifically: For each person in the candidate set, perform facial key point detection on the facial image data, extract geometric constraint points in the eye area, nose tip and mouth corner area, and determine whether there is an occlusion area based on the continuity of the distribution of the geometric constraint points; If there is an occluded area, the facial image data is input into a pre-trained multi-view pose estimation network, which outputs virtual facial texture maps from multiple different viewpoints. Select the corresponding pixel block from the virtual facial texture map that is closest to the current acquisition viewpoint, and fill it into the occluded area to form complete facial image data after completion. The completed facial image data is then re-correlated with the facial feature template using normalized cross-correlation to obtain an updated similarity score.

5. The intelligent inpatient registration and management method based on facial recognition according to claim 4, characterized in that, Based on the similarity score and the current department queue length for each candidate in the candidate set, an initial inpatient registration priority queue is generated, and the target person at the top of the initial inpatient registration priority queue is marked as pending registration. Specifically: The system obtains the number of people waiting for treatment and the average consultation time for each person in the candidate set who is currently registered in the hospital's real-time dispatch system, and calculates the expected waiting time for each person. The similarity scores of each person are sorted from high to low in the first sorting, and the expected waiting times are sorted from short to long in the second sorting; Assign a first weight value and a second weight value to the first ranking and the second ranking respectively, and then sum them up to obtain the comprehensive priority score for each person. The initial inpatient registration priority queue is formed by arranging the comprehensive priority scores from largest to smallest, and the target personnel at the top of the queue are identified as those awaiting activation of registration.

6. The intelligent inpatient registration and management method based on facial recognition according to claim 5, characterized in that, The first weight value and the second weight value are dynamically adjusted according to the overall congestion level of the hospital. The higher the congestion level, the larger the second weight value.

7. The intelligent inpatient registration and management method based on facial recognition according to claim 5, characterized in that, After marking the top-ranked target individual in the initial inpatient registration priority queue as having a pending registration status, the system further includes a bed pre-allocation module for allocating inpatient beds to the target individuals in the pending registration status, specifically: The list of currently available beds is read from the hospital bed management system. The list of available beds includes the floor location, room orientation, and care level attributes of each bed. Obtain the inpatient medical order information of the target personnel, which includes the required level of care and isolation requirements; The required level of care and isolation requirements are matched with the level of care and room orientation of each bed in the list of available beds to filter out a set of candidate beds that meet the matching conditions. Select the bed whose floor location is closest to the department of the target person from the candidate bed set, and use it as the pre-assigned bed.

8. The intelligent inpatient registration and management method based on facial recognition according to claim 7, characterized in that, After marking the target person as pending registration and completing bed pre-allocation, the system also includes a movement trajectory tracking module, used to monitor the movement trajectory of the target person's face in real time and determine whether the target person has actually arrived at the floor where the pre-allocated bed is located. Specifically: The sequence of facial image frames of the target person is continuously captured by at least three fixed-point camera devices deployed on the floor where the pre-assigned bed is located, and the offset of the pupil center coordinates of the target person between adjacent image frames is calculated. Based on the pupil center coordinate offset and the calibration parameters of the fixed-point camera device, the real-time spatial coordinates of the target person on the floor plan are inferred. The real-time spatial coordinates are compared with the bed coordinates corresponding to the pre-assigned bed using Euclidean distance. When the Euclidean distance is less than the preset arrival determination distance, it is confirmed that the target person has actually arrived at the pre-assigned bed.

9. The intelligent inpatient registration and management method based on facial recognition according to claim 8, characterized in that, After confirming that the target person has actually arrived at the pre-assigned bed, the system also includes an automatic registration information filling module, which is used to initiate the automatic filling process of hospital registration information, specifically: Retrieve the target person's historical basic personal information from the electronic health record database. The historical basic personal information includes name, gender, date of birth, allergic drugs, and past medical history. The current system time is obtained as the registration time, and the bed number of the pre-assigned bed, the registration time, and the historical personal basic information are combined into an initial registration electronic form; The hospital registration information verification rule library is invoked. The verification rule library contains rules for the integrity of required fields and the legality of data formats. The initial registration electronic form is verified item by item to generate a formal hospital registration record that passes the verification.

10. The intelligent inpatient registration and management system based on facial recognition according to claim 9, characterized in that, After generating the formal hospitalization registration record, the system further includes an encrypted storage and voucher generation module, used to encrypt and store the formal hospitalization registration record and generate an electronic hospitalization voucher bound to the facial feature template of the target person, specifically: A hash algorithm is used to perform a digest operation on the personnel identification field in the formal hospitalization registration record to obtain a fixed-length digest value as an encrypted index key; The formal hospitalization registration record is stored in the blockchain node according to the encrypted index key, and the encrypted index key is associated and bound with the facial feature template of the target person; An electronic hospitalization certificate in the form of a two-dimensional graphic code is generated based on the encrypted index key, and the electronic hospitalization certificate is pushed to the target person's mobile terminal.

Images