Medical fabric RFID identification management method and system based on visual detection and computer device

Abstract

The present application relates to the technical field of medical fabric management, and specifically provides a medical fabric RFID identification management method and system based on visual detection and computer equipment, the method comprising: placing the medical fabric in a preset detection area; identifying the fabric category, quantity and appearance state through a fusion model, and outputting the corresponding confidence, generating a visual data set; reading the RFID tag through an RFID read-write unit and generating a radio frequency data set; associating and matching the visual data set with the radio frequency data set, performing weighted judgment according to the confidence, and generating a judgment result; marking the abnormal attribution type according to the judgment result, constructing a fabric state matrix with the identity as the index, and updating the full life cycle record. The present application effectively overcomes the problem of missed reading caused by fabric stacking and metal interference, and solves the defects of difficult abnormal attribution and opaque flow nodes in the traditional scheme.

Description

Technical Field

[0001] This invention relates to the field of medical fabric management technology, specifically to a method, system, and computer equipment for RFID identification and management of medical fabrics based on visual inspection. Background Technology

[0002] Medical textiles (such as surgical gowns, bed sheets, and work clothes) come into direct contact with patients and medical staff in the daily operation of medical institutions. The quality of their cleaning and disinfection and the efficiency of their circulation are directly related to infection control and the safety of medical services. Therefore, they must be strictly managed.

[0003] Currently, laundry facilities and hospital logistics generally use a combination of manual counting and barcode scanning or single RFID reading for management. However, this approach faces significant challenges in practical applications. Fabrics often enter the inspection area in stacked or bundled form, severely obstructing visual recognition. Furthermore, RFID tags are often shielded by metal carts and subject to collisions between multiple tags, resulting in numerous missed reads. Traditional solutions rely solely on whether an electronic tag is read, lacking signal quality assessment, leading to extremely high false positive rates in scenarios with strong interference, such as motor start-stop, metallic environments, and wet fabrics. In addition, existing visual perception and RFID data are independent, making it impossible to simultaneously acquire the fabric's appearance (e.g., bloodstains, damage), hindering timely risk identification. When anomalies occur, it's difficult to accurately pinpoint the cause and the specific point of occurrence, only triggering general alarms and failing to achieve full lifecycle traceability management.

[0004] To address these issues, the present invention provides a visual inspection-based RFID identification and management method, system, and computer equipment for medical fabrics, in order to resolve the aforementioned problems. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a method, system and computer equipment for RFID identification and management of medical fabrics based on visual inspection, so as to solve the problems in the prior art.

[0006] One embodiment of the present invention provides a method for RFID identification and management of medical fabrics based on visual inspection, comprising the following steps: The medical fabric to be processed is placed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. The visual detection unit of the detection device identifies the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model, and outputs the category confidence, quantity confidence, and appearance confidence respectively to generate a visual recognition dataset. The RFID reading and writing unit of the detection device reads the RFID tags embedded in the TCM fabric in the preset detection area with radio frequency carrier signals within a single detection cycle. The number of successful reads per unit time is counted and the instantaneous field strength value of the reflected signal is extracted to generate a radio frequency dataset including unique department information and identity identifier. The readability confidence level of each identity identifier is evaluated based on the number of successful reads and the instantaneous field strength value. The visual recognition dataset and the radio frequency dataset are associated and matched. Weighted judgment is performed according to category confidence, quantity confidence, appearance status confidence and readability confidence level. Based on the judgment results, the abnormal attribution type of each medical fabric is determined. Using the identity identifier as an index, the current circulation node and abnormal attribution type are associated with the corresponding identity identifier to construct the fabric status matrix. The fabric lifecycle record in the background management system is updated based on the aforementioned transfer node.

[0007] This application also relates to a vision-based RFID identification and management system for medical fabrics, comprising: The fabric positioning module is used to place the medical fabric to be processed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. The visual recognition module is used to identify the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model using the visual detection unit of the detection device. It outputs the category confidence, quantity confidence, and appearance confidence respectively, and generates a visual recognition dataset. The radio frequency acquisition module is used to read the RFID tags embedded in the TCM fabric in the preset detection area with radio frequency carrier signals within a single detection cycle using the RFID reading and writing unit of the detection device. It counts the number of successful reads per unit time and extracts the instantaneous field strength value of the reflected signal to generate a radio frequency dataset including unique department information and identity identifier. Based on the number of successful reads and the instantaneous field strength value, it evaluates the readability confidence level of each identity identifier. The data processing module is used to associate and match the visual recognition dataset and the radio frequency dataset, perform weighted judgment according to category confidence, quantity confidence, appearance status confidence and readability confidence level, determine the abnormal attribution type of each medical fabric based on the judgment result, and use the identity identifier as an index to associate the current flow node and abnormal attribution type with the corresponding identity identifier to construct the fabric status matrix. The periodic update module is used to update the fabric lifecycle records in the background management system based on the flow nodes.

[0008] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described visual inspection-based RFID identification and management method for medical fabrics.

[0009] The above embodiments provide a visual inspection-based RFID identification and management method, system, and computer equipment for medical fabrics, which have the following beneficial effects: This invention eliminates the risks of stacking obstruction and multi-tag collisions by placing medical fabric in a single-layer flat position within the overlapping detection area of ​​the visual and RFID reading / writing space, ensuring consistent data acquisition. A pre-built multi-task branch fusion model synchronously identifies the fabric's category, quantity, and appearance, outputting multiple confidence levels, enabling the perception of fabric attributes and the detection of risks such as bloodstains and damage. Quantitative acquisition and hierarchical evaluation of the number of successful RFID tag reads and instantaneous field strength values ​​per unit time overcomes signal fluctuations caused by metallic environments, motor interference, and wet fabrics, improving the anti-interference capability and reliability of RFID. By associating and matching the visual and RFID datasets and performing weighted judgments, the counting and attribute benchmarks can be dynamically switched according to different confidence levels. Finally, a fabric state matrix containing anomaly attribution types and flow nodes is constructed using the identification identifier as an index, and the entire lifecycle record is updated in real time, solving the shortcomings of traditional solutions such as difficulty in anomaly attribution and opaque flow nodes. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a vision-based RFID identification and management method for medical fabrics, provided as an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0012] Reference Figure 1 One embodiment of the present invention provides a method for RFID identification and management of medical fabrics based on visual inspection, comprising the following steps: S10. The medical fabric to be processed is placed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. S20. Using the visual detection unit of the detection device, identify the category, quantity and appearance of TCM fabrics in the preset detection area based on the pre-built multi-task branch fusion model, and output the category confidence, quantity confidence and appearance confidence respectively to generate a visual recognition dataset. S30. Using the RFID reading and writing unit of the detection device, the RFID tags embedded in the TCM fabric in the preset detection area are read with radio frequency carrier signals within a single detection cycle. The number of successful reads per unit time is counted and the instantaneous field strength value of the reflected signal is extracted to generate a radio frequency dataset including unique department information and identity identifier. Based on the number of successful reads and the instantaneous field strength value, the readability confidence level of each identity identifier is evaluated. S40. The visual recognition dataset and the radio frequency dataset are associated and matched. Weighted judgment is performed according to category confidence, quantity confidence, appearance status confidence and readability confidence level. The abnormal attribution type of each medical fabric is determined according to the judgment result. The current circulation node and abnormal attribution type are associated with the corresponding identity identifier using the identity identifier as an index to construct the fabric status matrix. S50. Update the fabric lifecycle record in the background management system based on the flow node.

[0013] In this embodiment, as described in steps S10~S50 above, specifically: Step S10 is used to arrange the medical fabrics to be processed before testing. The testing device is a special testing equipment for medical fabrics that integrates a visual inspection unit and an RFID reading and writing unit. The preset testing area is a fixed workstation area marked on the equipment where the visual acquisition area and the RFID reading and writing area overlap. By placing the medical fabrics to be processed in a single-layer flat position in this overlapping workstation area, the occlusion problem caused by fabric stacking and bundling can be avoided. This prevents visual imaging from being blocked and affecting the acquisition of fabric type, quantity, and appearance status. It also reduces radio frequency signal collisions and shielding misses caused by multiple tag stacking. The unified acquisition space for visual and radio frequency data makes the residence time of a single fabric in this area constitute the sampling window for subsequent single testing cycles, providing a basis for the correlation and matching of visual and radio frequency data.

[0014] Step S20 is used to perform multi-dimensional visual recognition processing on medical fabrics placed within a preset detection area. The visual detection unit is an image acquisition and processing module (such as an industrial camera and edge computing box) mounted on the detection device. The pre-built multi-task branch fusion model is a neural network model deployed on the visual detection unit, capable of simultaneously completing multiple recognition tasks. The visual detection unit calls this multi-task branch fusion model to simultaneously identify the fabric category, quantity, and appearance status, and simultaneously outputs the corresponding three types of confidence scores. The model outputs category-aware features, location-aware features, and anomaly-aware features through a shared feature extraction layer, which are then input into the category recognition branch, quantity statistics branch, and appearance status analysis branch for processing. Finally, the fabric category, quantity, appearance status, and their respective confidence scores are encapsulated into structured data entries and integrated to generate a standardized visual recognition dataset. This avoids the time-consuming and lengthy process of using multiple single-task models for separate inference, and the inability to simultaneously output recognition confidence scores. It completely retains the corresponding features and recognition reliability data of the fabric's visual aspect, providing a complete visual judgment basis for subsequent weighted judgment with the radio frequency dataset.

[0015] Step S30 is used to perform radio frequency signal acquisition and tag credibility assessment on medical fabrics within a preset detection area. The RFID reading / writing unit is a radio frequency signal transmission and analysis module mounted on the detection device. A single detection cycle refers to the complete time window during which a single piece of fabric remains within the overlapping detection area to complete simultaneous visual and radio frequency acquisition. The radio frequency carrier signal is the radio frequency excitation signal emitted outward by the RFID reading / writing unit to wake up the fabric tag. The instantaneous field strength value is the radio frequency signal strength value corresponding to the tag's feedback reflection signal at the acquisition moment. The RFID tag embedded in the medical fabric pre-stores unique department information and identification information corresponding to the fabric. The identification information is an independent code that distinguishes a single piece of medical fabric. By utilizing the RFID reading / writing unit in the single unified detection corresponding to S10 and S20... During the testing period, radio frequency carrier signals are transmitted to read RFID tags on all medical fabrics within the area in batches. The number of successful reads of each tag per unit time is counted simultaneously, and the instantaneous field strength value of the reflected signal returned by the tag is extracted. A standardized radio frequency dataset containing department information and identification is generated. The readability confidence level of the RFID tag corresponding to each identification is comprehensively evaluated by combining the number of reads and the instantaneous field strength value. This avoids the shortcomings of traditional RFID detection, which relies solely on whether the read is successful as a single criterion and does not evaluate signal quality and reading stability. It also avoids the high false judgment rate in scenarios with metal shielding, electromagnetic interference, and wet fabrics. This enables the quantitative grading of RFID tag reading quality and provides an accurate radio frequency confidence basis for the subsequent weighted fusion judgment of visual datasets and radio frequency datasets.

[0016] Step S40 is used to perform data fusion, weighted judgment, and status archiving on the generated visual recognition dataset and RFID dataset. Specifically, association matching establishes a one-to-one binding relationship between visual recognition data and RFID tag radio frequency data within the same detection cycle and area; weighted judgment is a judgment method that comprehensively outputs the recognition result by combining multi-dimensional confidence quantification indicators; anomaly attribution type is the specific problem category classified for detection anomalies; flow node is the current detection process position of the medical fabric throughout its entire life cycle; and the fabric status matrix is ​​a structured status summary data that integrates the individual status, anomaly type, and flow position of all fabrics. This is achieved by performing one-to-one association matching between the visual recognition dataset and the RFID dataset, combined with the fabric category confidence level. The system employs a multi-dimensional weighted comprehensive assessment, considering factors such as quantity confidence, appearance confidence, and RFID tag readability confidence. Based on the weighted assessment results, it distinguishes and determines the anomaly attribution type for each medical fabric. Simultaneously, using the fabric's unique identifier as a retrieval index, it binds and associates the current circulation node and the determined anomaly attribution type to the corresponding fabric, thus uniformly integrating the detection status information of all fabrics to construct a fabric status matrix. This approach avoids the shortcomings of traditional solutions where visual and RFID data are independent and cannot be integrated for verification, and anomalies can only generate general alarms without accurately locating the problem type and occurrence node. It achieves precise attribution of detection anomalies and accurate recording of circulation nodes, providing standardized and structured status data support for subsequent full lifecycle traceability management of fabrics.

[0017] Step S50 is used to archive and update the fabric status matrix after the inspection is completed. The flow node is the current business process position of each fabric marked after S40 (such as RFID counting completed, visual verification completed, awaiting manual re-inspection, equipment abnormality). The background management system is a fabric management platform deployed in the cloud or on a local server. The full life cycle record is the data on the status changes of each business link from the incoming count to the outbound delivery of the fabric, with the identity identifier as the primary key. By retrieving the fabric status matrix constructed in S40, the current flow node and abnormal attribution type corresponding to each identity identifier in the matrix are extracted. The inspection result is written into the corresponding full life cycle record of the fabric in the background management system with the identity identifier as the retrieval primary key, thus completing the status archiving of this batch of fabrics. This can avoid the defects of traditional solutions that rely on paper ledgers or discrete electronic spreadsheet records, have gaps in flow nodes, and have no traceability of the historical status of individual fabrics. It realizes the status closed loop of medical fabrics from the incoming, washing, disinfection, outbound delivery to recycling links, and provides a traceable data foundation for hospital infection control and refined operation management.

[0018] In one feasible embodiment, assuming a medical institution recycles a batch of medical surgical gowns and bed sheets to be tested as fabrics to be processed, the entire process is completed based on an integrated vision and RFID detection device. The detection device is a tabletop linen sorting operation table structure. The tabletop is divided into workstations where the vision acquisition area and the RFID reading and writing area completely overlap. An industrial camera is mounted above the tabletop to form a vision detection unit. Multiple sets of ultra-high frequency RFID reading and writing antennas are preset inside the tabletop to form an RFID reading and writing unit. An edge computing box is mounted on the edge of the device to deploy a multi-task branch fusion model and radio frequency data processing program. Those skilled in the art can choose an existing mature linen RFID sorting operation table on the market as the basic hardware carrier. For example, a high-definition industrial camera can be added and the multi-task fusion algorithm of this invention can be deployed to adapt to the use of this method. This is clear and reproducible to those skilled in the art.

[0019] The staff first performed step S10, laying each piece of medical fabric in a single layer in the overlapping work area to avoid stacking obstruction and label signal interference, and to unify the spatial benchmark for dual-type data collection. Then, step S20 is executed, where the visual detection unit calls the pre-trained multi-task branch fusion model to complete image reasoning, and synchronously outputs the category, quantity, damage / bloodstain appearance status and corresponding confidence of each piece of fabric, and encapsulates it to generate a structured visual recognition dataset. Next, step S30 is executed. The RFID reading and writing unit transmits radio frequency carrier signals in the same single detection cycle, reads the high-temperature resistant washing RFID tags embedded in the fabric, counts the number of reads per unit time, extracts the instantaneous field strength value of the reflection, generates a radio frequency dataset containing department information and unique identification, and completes the classification of the readability confidence level of each tag based on the number of reads and field strength. Then, in step S40, the system associates and matches the visual recognition dataset and the radio frequency dataset of the same period, and makes a weighted judgment by combining four types of confidence: category confidence, quantity confidence, appearance status confidence and readability confidence level. It distinguishes three attribution types: normal recognition, abnormal visual recognition, and abnormal RFID tag, and simultaneously marks the current fabric flow node, and constructs a fabric status matrix with the identity identifier as the index. Finally, in step S50, the system extracts all fabric flow nodes and anomaly attribution information from the fabric status matrix, synchronously writes them into the cloud-based backend management system, updates the full lifecycle time sequence record of each fabric linked by its identity identifier, completely preserves the data of this inspection process, and realizes the closed-loop traceability of medical fabrics from entry inspection, anomaly judgment to flow archiving, solving the defects of traditional manual counting, single RFID reading mode with high misjudgment, data fragmentation, and inability to accurately trace anomalies.

[0020] In one embodiment, step S20, the construction of the pre-built multi-task branch fusion model, specifically includes: S201. Construct a medical fabric sample set, which includes several images of medical fabrics in different appearance states, and each medical fabric image is labeled with a fabric category label, quantity bounding box, appearance state label and feature point label. S202. Construct a multi-branch model architecture, which includes a shared feature extraction layer and parallel-connected category recognition branch, quantity statistics branch and appearance state analysis branch; S203. Construct a joint loss function, which is composed of a weighted sum of a cross-entropy loss term for constraining the category recognition branch, a smoothing absolute error loss term for constraining the quantity statistics branch, a focusing loss term for constraining the appearance state analysis branch, and a reprojection loss term for constraining the feature point labels. S204. Divide the medical fabric sample set into a training set and a validation set. Based on the joint loss function, iteratively train the multi-branch model architecture using the training set and validate it using the validation set. When the validation set loss value converges, the trained multi-task branch fusion model is obtained.

[0021] In this embodiment, as described in steps S201~S204 above, specifically: Step S201 is used to complete the construction of the basic dataset required for training the multi-task branch fusion model and the annotation processing of the training data. The medical fabric sample set is a collection of image data collected from real laundry plants or medical institutions, covering various fabric morphologies and working conditions, used for supervised learning training of the model. Different appearance states include various actual working conditions such as intact medical fabrics, partial damage, stains, wrinkles, and deformation. Fabric category labels are used to identify specific fabric types in the images, such as surgical gowns, bed sheets, work clothes, and drapes. The quantity bounding box refers to the rectangular annotation box drawn for each independent fabric in the image. The labels are: Box (for identifying the location and quantity of each item), Appearance status labels (for marking whether there are any defects such as tears, bloodstains, or other contamination on the surface of a single piece of fabric), and Feature point labels (for marking physical points with fixed geometric significance on key parts of the fabric, such as the four corners of a bed sheet, the center point of the collar of a surgical gown, the center points of the left and right cuffs, and the center point of the hem). By collecting several images of medical fabrics covering different lighting conditions, different placement densities, and different degrees of deformation, and having each image simultaneously assigned the above four types of labels by annotators or automated annotation tools, a standardized multi-dimensional labeled sample set is constructed. This avoids the shortcomings of traditional visual sample sets, which only have single classification labels, lack quantity positioning and defect feature labels, and cannot adapt to multi-task synchronous training. It provides sufficient and standardized sample data support for subsequent multi-branch model category recognition, quantity statistics, appearance anomaly analysis, and feature matching training, ensuring the accuracy and generalization ability of the model in multi-task learning.

[0022] Step S202 is used to complete the overall network architecture construction and branch structure configuration of the multi-task branch fusion model. The shared feature extraction layer is a visual feature encoder composed of a Backbone in a convolutional neural network (such as ResNet-50 or MobileNet-V3; when deployed on an edge computing box, both accuracy and computing power must be considered) combined with a Feature Pyramid Network (FPN). This layer is used to extract multi-scale fusion features with both semantic and spatial information from the input medical fabric image. This layer is shared by all subsequent branches to avoid computational redundancy caused by repeated feature extraction in each task. The category recognition branch, quantity statistics branch, and appearance state analysis branch are three independent subnetworks connected in parallel to the output of the shared feature extraction layer: the category recognition branch consists of a global pooling layer and a fully connected layer, used to output the probability distribution and category confidence of each fabric type; the quantity statistics branch consists of a convolutional regression head and an anchor box matching mechanism. The first branch outputs the bounding box coordinates and quantity confidence of each independent fabric in each image, and also regresses the coordinates of key feature points of the fabric (such as the corners of a bed sheet, the collar and cuffs of a surgical gown) to connect with the subsequent reprojection loss constraints. The appearance state analysis branch consists of a local attention convolution head, which focuses on the surface area of ​​the fabric and outputs the appearance state classification and appearance state confidence. By connecting the above three branches in parallel to the multi-scale output of the shared feature extraction layer, the same fabric image can simultaneously output the three categories of results (category, quantity, and appearance state) and their corresponding confidence in a single forward inference. This avoids the time-consuming redundancy caused by serial inference in single-task models, the redundancy of independent feature extraction in multiple tasks, and the technical problems that a single model cannot simultaneously take into account classification accuracy and localization accuracy. It improves the model's computational efficiency while ensuring the accuracy of each recognition task, and provides a stable and standardized network architecture foundation for training under joint loss function constraints and realizing high-precision fusion recognition of multiple tasks.

[0023] Step S203 is used to configure a loss constraint mechanism for multi-task collaborative optimization for the multi-branch network architecture built in S202, so that the shared feature extraction layer and the three branches can take into account classification accuracy, localization accuracy and anomaly detection capability during the training phase; wherein, the joint loss function refers to the composite loss obtained by weighting and summing the loss terms of multiple tasks according to the importance of the tasks.

[0024] The cross-entropy loss term is used to constrain the category recognition branch, measuring the difference between the predicted fabric type probability distribution and the true category label, driving the learning of category-aware features; its standard calculation formula is: in, The total number of fabric categories, This is a unique hot tag for the real category. Predict the output probability of the corresponding category for the model.

[0025] The Smooth L1 Loss term is used to constrain the bounding box regression and feature point coordinate regression of the statistical branch. Compared with MSE, it is better for outlier boxes and adapts to bounding box annotation noise caused by fabric wrinkles and stacked edges. Its standard calculation formula is: in, This refers to the deviation between the predicted coordinates and the actual labeled coordinates.

[0026] The Focal Loss term is used to constrain the appearance condition analysis branch. It addresses the class imbalance problem in medical fabric appearance defect detection by reducing the weight of simple negative samples (i.e., fabrics with normal appearance) and increasing the weight of difficult-to-distinguish positive samples (i.e., fabrics with defects such as tears and bloodstains, typically accounting for less than 5% of the samples). Its standard calculation formula is: in, This is a modulation factor used to adjust the weight difference between easy and difficult samples.

[0027] The reprojection loss term constrains the feature point labels, calculating the Euclidean distance or smoothed L1 distance between the regressed feature point coordinates and the labeled keypoint coordinates in the image space or normalized space. This forces the shared layers to learn the prior geometric structure of the fabric, offsetting the keypoint drift caused by the soft material and suspension / laying deformation of medical fabrics. This scheme uses the smoothed absolute error form, and the calculation formula is as follows: in, To predict the coordinates of feature points, Provide the actual coordinates of the feature points.

[0028] By assigning the above four losses to preset weighting coefficients (such as...) =1.0, =1.0, =2.0, =0.5, where Focal Loss =2、 =0.25 is the standard setting) Weighted summation yields the joint loss: Backpropagation updates network parameters during training iterations; this avoids the problem that a single loss function cannot meet the needs of multiple tasks, or that simple summation of multiple tasks can lead to a dominant gradient in one branch causing underfitting in other branches. It ensures that the four tasks of category recognition, quantity statistics, appearance anomaly analysis and feature point regression converge collaboratively on a shared skeleton, providing training assurance for the output of high-confidence category, quantity, appearance status and three-way perception features in the S20 inference stage.

[0029] Step S204 is used to complete the supervised training, generalization verification, and early stopping judgment of the multi-task branch fusion model. The labeled sample set constructed in S201 and the multi-task joint loss function configured in S203 are input into the multi-branch network architecture built in S202 to complete the parameter iterative solution, resulting in an edge-deployable inference model. The training set refers to a subset of samples from the medical fabric sample set divided in a fixed ratio (e.g., 8:2 or 7:3) to participate in gradient backpropagation and update network weight parameters. This subset is used by the model to learn fabric categories, quantity boundaries, and appearance defects. The core data source for the trap is the geometric feature points; the validation set is an independently reserved subset of samples that does not participate in parameter updates and is only used to evaluate the model's generalization ability after each iteration; convergence refers to the steady-state state of the model where the joint loss value of the validation set tends to stabilize with the number of iterations and no longer decreases significantly. In engineering, this can be determined by a preset threshold (e.g., the loss decrease is less than 0.001 after 10 consecutive training iterations) to avoid overfitting caused by excessive iteration; the medical fabric sample set is divided into training and validation sets according to a preset ratio, and the joint loss function is constructed based on S203. The multi-branch model architecture of S202 is trained iteratively multiple times using the training set, continuously optimizing the network parameters of the shared feature extraction layer and the three task branches. After each training round, lossless evaluation is performed on the validation set, and the convergence trend of the joint loss of the multi-task is monitored in real time. When the loss value of the validation set meets the convergence judgment condition, the training is terminated in time, the optimal network weights are fixed, and the trained multi-task branch fusion model is obtained. This model can be exported in ONNX or TensorRT format and deployed to the edge computing box of the detection device for real-time inference calls in the S20 step. This can avoid training defects such as overfitting caused by dataset mixing and underconvergence caused by premature termination of training. It solves the problems of asynchronous convergence of the branches of the multi-task model and weak generalization ability in real working conditions. It ensures that the model can simultaneously take into account the comprehensive recognition accuracy of classification, localization, defect detection and feature point regression in scenarios such as complex deformation of medical fabrics, weak light interference and sparse defects. It provides mature and reliable model computing power support for the stable and high-precision output of standardized visual recognition datasets in the subsequent S20 step.

[0030] In one embodiment, step S20 specifically includes: S21. Use the visual inspection unit of the inspection device to acquire images of medical fabrics in a preset inspection area; S22. Input the medical fabric image into the multi-task branch fusion model, and perform feature extraction and fusion on the medical fabric image through the shared feature extraction layer to obtain fused features including category-aware features, location-aware features and anomaly-aware features; S23. Input the fused features into the category recognition branch, the quantity statistics branch, and the appearance state analysis branch respectively. The category recognition branch performs classification processing on the category-aware features and outputs the fabric category and the corresponding category confidence score. The quantity statistics branch performs bounding box regression processing on the position-aware features and outputs the fabric quantity and the corresponding quantity confidence score. The appearance state analysis branch performs anomaly classification processing on the anomaly-aware features and outputs the appearance state and the corresponding appearance state confidence score. S24. Encapsulate the fabric type, fabric quantity, appearance status and their corresponding confidence scores into structured data entries to generate a visual recognition dataset.

[0031] In this embodiment, as described in steps S21-S24 above, specifically: Step S21 is used to acquire the original image of the medical fabric under single-frame detection conditions, providing standardized image input for feature extraction and multi-dimensional recognition of the multi-task branch fusion model. The visual inspection unit is a high-definition industrial camera and image acquisition module mounted on top of the detection device. It can match the spatial size of the preset detection area to complete fixed-field imaging, supports constant exposure, automatic white balance, and anti-reflective imaging parameter configuration, and is adaptable to complex conditions such as the white background of medical fabrics, smooth and reflective materials, and fluctuating ambient light. The preset detection area is the area where the visual acquisition area and the RFID reading / writing area overlap in S10. With a fixed workstation area, the field of view is fully aligned with the overlapping workstations, ensuring that each frame of image contains only the currently inspected flat medical fabric, avoiding redundant background interference. The visual detection unit collects images of the flat medical fabric at fixed points and with a fixed field of view, obtaining original fabric image data with clear texture, realistic deformation, and no stacking or occlusion. This avoids image defects and cluttered backgrounds caused by random snapshots and field of view shifts, ensuring that the images input to the model are consistent in terms of acquisition scale, imaging environment, and placement conditions, providing a regular and consistent original image input foundation for high-precision feature extraction and multi-task branch synchronous inference.

[0032] Step S22 is used to perform unified multi-scale feature extraction and feature fusion operations on the acquired original images of medical fabrics, generating composite fusion features that can be parsed in parallel by three task branches. The multi-task branch fusion model is a lightweight neural network model trained and deployed within an edge computing box. The shared feature extraction layer consists of a backbone network combined with a feature pyramid network (FPN), capable of simultaneously extracting deep semantic information and shallow spatial detail information of the image. Category-aware features are global semantic features characterizing the fabric's material, style, and category attributes, used to distinguish different fabric types such as surgical gowns, bed sheets, and linens. Position-aware features are local geometric features carrying spatial coordinate information such as fabric contours, edges, and corners, used to support bounding box regression and key feature point localization. Anomaly-aware features... To focus on fine-grained features of local texture differences in fabrics, this method captures appearance defects such as stains, tears, and wrinkles. Medical fabric images acquired by S21 are preprocessed and input into a multi-task branch fusion model. Multi-level feature extraction is performed through a shared feature extraction layer, and cross-layer fusion is conducted on features at different scales. Simultaneously, three independent but interconnected fused features—category-aware features, location-aware features, and anomaly-aware features—are decoupled and output. This avoids the repetitive computation and inference delays caused by extracting single-class features in stages. It simultaneously produces the features required for the three tasks based on a single forward operation, ensuring real-time inference speed for edge devices. Furthermore, by using a unified feature benchmark based on a shared backbone, it reduces feature offset between different tasks, providing a unified and complete fused feature input for different branches to simultaneously complete classification, localization, and defect identification.

[0033] Step S23 is used to complete the multi-branch split parsing and multi-task parallel reasoning processing of the fused features, so as to realize the synchronous recognition and reliable quantitative output of fabric category, quantity location and appearance status. Among them, the category recognition branch, quantity statistics branch and appearance status analysis branch are three parallel task subnets built in S202 and trained by the joint loss function constraint in S203. Each branch performs its own function, does not interfere with each other and has the same feature source, and completes the exclusive task reasoning based on the shared fused features. The category perception feature, location perception feature and anomaly perception feature are multi-dimensional fused feature information extracted in S22, which respectively correspond to the task recognition requirements of the three branches.

[0034] Specifically, the category recognition branch performs global semantic classification operations on category-aware features. Through feature weight matching and probability normalization, it determines the specific category of the current medical fabric and completes the inference output based on the cross-entropy loss constraint logic of the S203 training phase. The predicted probability of each category is obtained through softmax normalization, and the calculation logic is adapted to the cross-entropy loss function. The corresponding inference mapping relationship outputs the category confidence score of the corresponding product category, quantifying the reliability of category identification.

[0035] The statistical branch performs bounding box regression calculations based on location-aware features, fitting the contour edges and coordinate ranges of each independent fabric in the image, selecting single fabric targets, distinguishing multiple individual fabrics, and counting the number of valid targets. The coordinate regression accuracy of this branch is constrained by the smoothed absolute error loss during the training phase, and the fitting formula is used to... Simultaneously, combined with feature point reprojection loss Correct the offset of key point coordinates of fabric, output the number of fabrics and the corresponding confidence level, and quantify the accuracy of fabric counting and positioning.

[0036] The appearance condition analysis branch performs local fine-grained anomaly classification operations on anomaly perception features, focusing on fabric surface texture and integrity features to identify different appearance conditions such as intact, damaged, bloodstains, and contamination. The reasoning ability of this branch for difficult and imbalanced samples is supported by the optimization of the focusing loss function during the training phase, and the formula is adapted accordingly. The training optimization logic effectively improves the accuracy of defective fabric identification, outputs the corresponding appearance state confidence level, and quantifies the credibility of defect identification.

[0037] By splitting the unified fusion features into three independent branches, the three inference tasks of classification, localization, and defect detection are completed simultaneously, and the fabric category, fabric quantity, appearance status, and corresponding three confidence levels are output at once. This avoids the shortcomings of traditional step-by-step recognition schemes, such as task serialization, secondary feature extraction, and low inference efficiency. At the same time, based on the homogeneous fusion features and the previous multi-loss joint training mechanism, the recognition benchmark of multiple tasks is guaranteed to be unified, eliminating the recognition deviation of different tasks. This enables synchronous, accurate, and quantifiable recognition of the multi-dimensional status of medical fabrics, providing complete and accurate multi-dimensional recognition data support for the standardized generation of the subsequent S24 structured dataset.

[0038] Step S24 is used to complete the structured encapsulation and standardized dataset construction of the multi-task inference results, realizing the organization and traceable digital storage of visual recognition results. Among these, fabric category, fabric quantity, and appearance status are the core recognition results output by the three-branch parallel inference; category confidence, quantity confidence, and appearance status confidence are the reliability metrics of the corresponding recognition results, used to characterize the reliability of each piece of visual recognition information. Structured data entries are standardized data units organized according to fixed fields and formats, uniformly containing all visual recognition information within a single detection cycle. The visual recognition dataset is a regularized visual data source formed by summarizing multiple sets of structured data entries, which can be used for subsequent data association matching and weighted judgment. The results are obtained through summarizing and inference. The fabric category and category confidence level, fabric quantity and quantity confidence level, and appearance condition and appearance condition confidence level are encapsulated into independent structured data entries according to preset data field specifications. A set of structured entries is generated for each single detection cycle. After multiple batches of detections are summarized, a complete, orderly, and searchable visual recognition dataset is constructed. This avoids the shortcomings of traditional visual recognition, such as scattered output results, lack of confidence labeling, and inconsistent data formats, which lead to the inability to quantify and compare data and the difficulty in linking and integrating with radio frequency data. It provides a standardized data foundation with unified format, complete fields, and quantifiable confidence for the subsequent S40 steps to realize the association and matching of visual datasets and RFID radio frequency datasets and the multi-confidence weighted fusion judgment, ensuring the rigor and accuracy of multi-source data fusion traceability of the whole machine.

[0039] In one embodiment, step S30 specifically includes: S31. Using the RFID reading and writing unit of the detection device, a radio frequency carrier signal is transmitted in a preset detection area at a preset power, and the reflected signal returned by each RFID tag based on the radio frequency carrier signal is received. The tag data containing unique department information and identity identifier is obtained by parsing. S32. Within a single detection cycle, the number of successful reads of the RFID tags per unit time is counted according to the identity identifier group, and the instantaneous field strength value corresponding to each identity identifier in the reflected signal is extracted. S33. Calculate the reading stability feature based on the number of successful reads per unit time, and calculate the signal strength feature based on the instantaneous field strength value. Calculate the comprehensive perception score of the RFID tag based on the reading stability feature and the signal strength feature. S34. Based on the comprehensive perception score, the readability confidence level of the corresponding identity identifier is divided into a certain level, a doubtful level, and a missing level, generating an RF dataset containing unique department information, identity identifier, and readability confidence level.

[0040] In this embodiment, as described in steps S31-S34 above, specifically: Step S31 is used to complete the stable excitation of the radio frequency signal and the acquisition and analysis of the tag's raw data within the preset detection area, providing standardized raw radio frequency data for subsequent tag reading statistics and signal quality assessment. The RFID reading / writing unit is a built-in ultra-high frequency radio frequency reading / writing module and array antenna of the detection device, supporting adjustable power radio frequency carrier transmission and tag reflection signal analysis, suitable for close-range, high-precision detection scenarios of medical fabrics. The preset power is the radio frequency transmission power calibrated for the overlapping detection distance of this equipment (typically adjustable range 5~30dBm), which can be finely adjusted according to the fabric's wet / dry state, placement density, and on-site electromagnetic environment, ensuring that tags within the area are fully awakened while avoiding cross-area cross-reading and multi-tag signal collision interference caused by excessive power. The preset detection area is the visual and RFID spatial overlapping workstation defined in S10, with the radio frequency signal coverage matching the placement area of ​​a single-layer flat-laid fabric, achieving directional, small-range, precise radio frequency excitation. The radio frequency carrier signal conforms to ISO-18000-6C / EPC. The Gen2 standard ultra-high frequency excitation signal is used for the energy wake-up and data interaction of passive RFID tags; the reflected signal is the data signal carried by the RFID tag embedded in the medical fabric after being excited by the carrier wave; the department information is the classification information of the medical institution to which the fabric belongs and the corresponding washing department pre-stored in the tag; the identity identifier is a unique code that binds to each piece of medical fabric, used to realize the traceability and differentiation of individual fabrics; the RFID read and write unit transmits radio frequency carrier signals in a directional manner with a calibrated preset power to fully cover and activate all RFID tags embedded in the medical fabrics in the detection area, and receives the reflected signals returned by each tag, demodulates, decodes and denoises the signals, extracts the department affiliation information and unique identity identifier built into each tag, and completes the batch collection of raw radio frequency tag data; it can avoid the defects of traditional RFID detection power fixed adaptability, uneven signal coverage, cross-area misread or insufficient excitation missed reads, and ensure that all flat fabric tags can be stably woken up and effectively read within a single detection cycle, providing a complete, accurate and complete raw radio frequency data foundation for subsequent S32 read count statistics, field strength extraction and signal quality quantitative evaluation.

[0041] Step S32 is used to complete the statistical analysis of RFID tag reading behavior and the extraction and processing of radio frequency signal feature values ​​within a single detection cycle, providing quantitative raw indicators for subsequent multi-dimensional radio frequency feature calculations. A single detection cycle is a fixed time window that matches the visual image acquisition sequence and synchronously completes image capture and radio frequency signal acquisition, achieving temporal alignment between visual and radio frequency data. The identification identifier is the unique code of a single piece of fabric obtained in S31, serving as an independent index for group statistics. The number of successful reads per unit time refers to the effective number of times the same identification tag successfully demodulates and decodes per unit time within the detection cycle, used to intuitively reflect the stability of the communication link between the tag and the antenna. The instantaneous field strength value is the RSSI signal strength carried in the reflected signal, characterizing the current moment of the RFID tag's reflection. The signal power of the backscattered signal reaching the read / write antenna fluctuates due to fabric stacking, wrinkles, metallic interference, and distance. Using a unique identifier as the grouping basis, all valid tag data within a single detection cycle is categorized and summarized. The number of successful reads per unit time for each identifier is counted, and the instantaneous field strength value corresponding to the identifier is extracted from the reflected signal of each tag. A one-to-one correspondence table of original radio frequency characteristics is established, linking the identifier, read count, and instantaneous field strength. This avoids the problems of mixed, ungrouped multi-tag data and the inability to distinguish the signal quality of individual fabrics. It transforms the messy original radio frequency messages into quantifiable and independently calculable standardized indicators, providing regular and distinguishable quantitative basic data for S33 to calculate read stability characteristics, signal strength characteristics, and comprehensive perception scores.

[0042] Step S33 is used to perform feature mapping and fusion of the two types of radio frequency quantization indicators extracted in S32 to construct a comprehensive quantization score that can uniformly evaluate the communication quality of a single RFID tag. Among them, the read stability feature is a normalized value obtained by mapping the number of successful reads per unit time. The higher the number of reads, the more stable the communication link between the tag and the read / write antenna, and the lower the probability of signal collision and missed reads. The signal strength feature is a quantized value obtained by normalizing the instantaneous field strength value. The higher the field strength value, the lower the backscattering signal loss of the tag and the better the communication signal-to-noise ratio. The comprehensive perception score is a single quantization score obtained by weighting the fusion stability feature and the signal strength feature, which is used to intuitively characterize the RF of the fabric within the current detection period. The overall readable status of the ID tag is determined by: firstly, converting the number of successful reads per unit time into a fixed range of reading stability characteristics based on a preset mapping function; then, normalizing the signal strength characteristics after applying mean filtering and removing abnormal jump values ​​to multiple sets of instantaneous field strength values ​​corresponding to the same identity identifier; finally, configuring preset weight coefficients to weight and sum the two types of features to obtain the comprehensive perception score of the corresponding identity RFID tag; this avoids the shortcomings of relying solely on a single number of reads or a single signal strength to unilaterally determine the readable status of the tag, taking into account both communication stability and signal transmission quality, quantitatively distinguishing the RFID reliability of different fabric tags, and providing a continuous and comparable quantitative basis for S34 to divide multi-level readable confidence levels.

[0043] Step S34 is used to complete the label readability classification based on the comprehensive perception score output by S33, and to standardize and encapsulate radio frequency related information to construct a complete radio frequency dataset. The readability confidence level is a three-level label classification based on the comprehensive perception score range: Confident level indicates that the comprehensive perception score is higher than the high-score threshold, the label's radio frequency communication is stable, the signal strength is sufficient, and the recognition result is highly reliable; Suspicious level indicates that the comprehensive perception score is in the middle range, the label's reading frequency fluctuates or the signal attenuates, and the recognition result is uncertain; Missing level indicates that the comprehensive perception score is lower than the low-score threshold, the label fails to read multiple times, the signal is weak or even there is no reflected signal, and it is determined that the fabric is not present in the current detection area. The radio frequency dataset is a structured data source formed by summarizing multiple standardized radio frequency data entries, with each entry uniformly bound to department information and individual items. The system comprises three core fields: unique fabric identification, tiered readability confidence level, and three other categories. By pre-setting two-level scoring thresholds, the comprehensive perception scores corresponding to each identification are matched to three confidence levels: certainty, doubt, and missing. Subsequently, the unique department information, identification, and tiering results parsed from S31 are encapsulated into independent radio frequency (RF) data entries. A set of RF entries is generated in a single detection cycle, and multiple rounds of detection are summarized to form an ordered and searchable RF dataset. This avoids the shortcomings of simply determining presence or absence without quantifying the credibility of identification. By differentiating the reliability of RF data through three levels of confidence, a unified judgment standard for the confidence dimension of visual recognition data and RF data is achieved. This provides standardized RF data support with format specifications and confidence stratification for subsequent fusion and matching of visual and RF datasets and multi-source confidence weighted judgment.

[0044] In one embodiment, step S33 specifically includes: S331. Calculate the dispersion of the number of successful reads per unit time and the reading stability feature, wherein the reading stability feature is inversely proportional to the variance of the number of successful reads; S332. Within the sliding time window of a single detection cycle, the instantaneous field strength value is filtered to remove outlier spikes caused by electromagnetic interference, resulting in a smoothed field strength sequence. S333. Calculate signal strength features based on the smoothed field strength sequence. The signal strength features are the difference between the mean of the field strength sequence and the current environmental noise baseline. The current environmental noise baseline is the mean of the radio frequency noise floor collected by the detection device when no tag passes by or the initial noise floor in the first frame detection period. S334. Multiply the reading stability feature by the signal strength feature to obtain the comprehensive perception score of the RFID tag.

[0045] In this embodiment, as described in steps S331-S334 above, specifically: Step S331 is used to quantify the stability of the radio frequency communication link based on the fluctuation of the tag reading frequency within a single detection cycle, and outputs standardized reading stability features, providing the first quantitative input for subsequent comprehensive perception scoring calculation; wherein, the dispersion is characterized by the sample variance of the number of successful readings per unit time. The larger the variance value, the more drastic the fluctuation of the number of readings of the same identification tag in different sub-time periods, the more serious the impact of multi-tag signal collision, fabric occlusion, and radio frequency multipath interference, and the worse the communication stability; the smaller the variance value, the more consistent the number of readings in each time period, and the continuous and reliable tag communication link; the reading stability feature is inversely proportional to the variance of the number of successful readings, which can be mathematically expressed as: in, To pre-determine the normalization constant, The variance of the number of successful reads per unit time is the sample variance. The larger the variance, the lower the value of the stability feature, which intuitively reflects the reliability of tag communication. Based on the successful read count sequence obtained by grouping and statistically analyzing multiple groups of read counts per unit time according to S32 identity identifiers, the overall dispersion of the read counts in this group is calculated using the sample variance formula. Then, the read stability characteristics with regular value intervals are obtained through inverse proportional mapping. This avoids the defect that relying solely on the average read count as a single indicator cannot distinguish between "uniform high-frequency reads" and "fluctuating reads with high and low values". It accurately quantifies the jitter and missed read fluctuations in the tag communication process, objectively reflects the long-term stability capability of the radio frequency link, and provides a standardized quantitative indicator with a stable dimension for subsequent calculation of comprehensive perception scores based on fused signal strength features.

[0046] Step S332 is used to perform local time-domain noise reduction and pulse interference removal on the instantaneous field strength time-series data extracted in S32, eliminating RSSI spike distortion caused by instantaneous electromagnetic pulses in the field, and outputting a stable field strength sequence that can be used for mean calculation, providing a clean original signal without abnormal jumps for subsequent solution of signal strength characteristics; wherein, the single detection period is a fixed time window for synchronous acquisition of radio frequency and image, and all instantaneous field strength values ​​are time-series sampling points acquired in chronological order within this period; the sliding time window is a fixed-length local time-domain analysis window (the number of sampling points can be configured in the project, such as 5 to 15 consecutive field strength sampling points), and the window slides along the time sequence to traverse all instantaneous field strength data, realizing local statistical judgment rather than global one-size-fits-all processing; outlier spikes generated by electromagnetic interference are instantaneous extreme RSSI values ​​caused by the start-up and shutdown of metal equipment in the field, radio frequency co-frequency interference, and switching power supply pulses, and do not belong to the true value of the fabric label. Reflected signals can inflate or deflat the average field strength, distorting signal strength characteristic calculations. The filtering process employs a robust local anomaly detection mechanism. Within each sliding window, the median and absolute deviation of the field strength are statistically analyzed. Extreme peaks exceeding a preset deviation threshold are identified as outliers and removed. The remaining valid sampling points are then smoothed through interpolation to fill in the time gaps, generating a continuous, smooth field strength sequence without abrupt shifts. Local filtering within the sliding window removes transient interference outliers and filters out random impulse noise from damaging the field strength values, preserving the genuine, smooth signal fluctuations caused by fabric obstruction and flexible deformation. This avoids the significant shifts in statistical results caused by extreme interference peaks when directly using the original instantaneous field strength sequence to calculate the mean, which can underestimate or overestimate the tag's true signal transmission capability. This ensures that subsequent S333 signal strength characteristic calculations are based solely on the tag's actual reflected signal, improving the objectivity and accuracy of RF signal quality quantification indicators.

[0047] Step S333 is used to eliminate inherent environmental radio frequency noise interference based on the noise-reduced and smoothed field strength time series, and to calculate the signal strength characteristics that can truly reflect the effective echo capability of the tag. The smoothed field strength series is an RSSI sampling set that eliminates outlier spikes in electromagnetic interference and has smooth time series fluctuations, avoiding interference from instantaneous impulse noise on the mean statistics. The mean of the field strength series is the baseline signal amplitude after averaging all smoothed field strength sampling points for a single tag, representing the average power level of the tag echo signal. The current environmental noise baseline is the average radio frequency environmental noise floor collected when there are no fabric tags and no effective backscattered signals in the detection area. There are two ways to obtain this: one is to pre-collect and calibrate the noise floor when the equipment is unloaded; the other is to record the initial noise floor recorded when there are no tags in the first detection cycle of the detection process, used to offset the fixed background noise brought by the on-site equipment and the spatial electromagnetic environment. The calculation logic of the signal strength characteristics is the smoothed field strength mean minus the environmental noise baseline, mathematically expressed as: in, Signal strength characteristics, The mean of the smoothed field strength sequence. The current environmental noise baseline is used as the reference. This difference can intuitively represent the effective signal-to-noise ratio level of the tag echo signal exceeding the environmental noise. The larger the difference, the stronger the tag's effective communication signal. First, the arithmetic mean of the smooth field strength sequence corresponding to a single identity tag is calculated. Then, the difference operation is performed on the environmental noise baseline corresponding to the current scene. Finally, the normalized signal strength feature, which eliminates environmental interference, is obtained. This avoids the defect that the signal strength index is not comparable horizontally due to the difference in background noise in different sites and time periods when directly using the original field strength mean. It eliminates the calculation bias caused by the electromagnetic background of a fixed environment and accurately extracts the effective reflected signal strength generated by the medical fabric tag itself. This provides a clean, unified, and comparable quantitative indicator of the signal dimension for solving the comprehensive perception score by fusing and reading stability features.

[0048] Step S334 is used to complete the fusion calculation of the read stability feature and the signal strength feature to obtain a unified quantitative comprehensive perception score of the RFID tag's radio frequency communication quality, which serves as the core criterion for classifying the readability confidence level in S34. The read stability feature is a quantitative index obtained in S331 based on the inverse solution of the variance of the number of reads, characterizing the magnitude of the timing fluctuation in tag communication. The signal strength feature is the effective echo signal-to-noise ratio index obtained in S333 after subtracting the environmental noise baseline, characterizing the effective amplitude level of the tag's reflected signal. The comprehensive perception score simultaneously couples the two core dimensions of link stability and effective signal strength, using a multiplication fusion method, and the calculation formula is as follows: When both the reading stability feature and the signal strength feature are at high values, the overall perception score is significantly higher, indicating stable tag reading and sufficient effective echo signal. If either indicator is low, it will directly lower the overall perception score, and can simultaneously identify two types of poor tag conditions: "stable reading but weak signal" and "strong signal but frequent reading jitter". By directly multiplying the reading stability feature output by S331 with the signal strength feature output by S333, the overall perception score of the corresponding identification RFID tag is obtained. This avoids the problem of a single indicator value being too high and masking the defects of another dimension when using weighted summation and fusion. The multiplication coupling mechanism has a strong suppression effect on the degradation of any radio frequency performance, and can objectively and evenly evaluate the overall radio frequency readability performance of a single fabric tag, providing an accurate and sensitive quantitative scoring standard for S34 to distinguish between the three levels of readability confidence: sure, doubtful, and missing.

[0049] In one embodiment, step S40, which involves associating and matching the visual recognition dataset with the radio frequency dataset, and performing a weighted determination based on category confidence, quantity confidence, appearance status confidence, and readability confidence level, specifically includes: S41. Extract the number of fabrics in the visual recognition dataset and the number of deduplicated reads of the identity tags in the radio frequency dataset, and compare the number of fabrics with the number of deduplicated reads, wherein the number of deduplicated reads is the number of RFID tags after deduplication of identity tags by the number of successful reads per unit time. S42. When the number of fabrics matches the number of deduplication reads, the association matching is determined to be successful. Based on the co-occurrence relationship in the preset detection area, a one-to-one correspondence between the fabrics and identity identifiers corresponding to each structured data entry in the visual recognition dataset is established. S43. Based on the association matching relationship, obtain the category confidence level, quantity confidence level, appearance status confidence level, and readability confidence level of the associated identity identifier corresponding to the associated fabric, and perform a weighted judgment with the comprehensive perception score, and output the judgment result: When the readability confidence level is at the certainty level and the comprehensive perception score is greater than the first preset threshold, it is determined that the identity identifier is used as the fabric counting benchmark, and the category and appearance status are used as the attribute correction benchmark, and the first determination result is output. When the readability confidence level is doubtful or missing, the quantity and appearance of the fabric are used as the judgment criteria, and a second judgment result is output. When the number of fabrics is inconsistent with the number of deduplicated reads, or when the readability confidence level is missing and the appearance status confidence level is lower than the second preset threshold, the association matching is determined to be unsuccessful and a tag anomaly alarm is triggered, and a third determination result is output.

[0050] In this embodiment, as described in steps S41-S43 above, specifically: Step S41 is used to complete the basic quantity alignment verification between the visual data source and the radio frequency (RF) data source. By comparing the quantities from both sources, it provides a preliminary verification basis for subsequent entity association matching and anomaly detection. Specifically, the visual recognition dataset is a structured dataset generated in S24 that completely stores the visual recognition information of the fabric within a single detection cycle. The fabric quantity is the total number of valid single fabric targets within the current detection area, calculated by the quantity statistics branch from the multi-task branch inference output in S23. The RF dataset is a standardized dataset generated in S34 that carries RFID tag RF identification information. The identification identifier is a unique code bound to a single piece of medical fabric and is the core index for RF data grouping and statistics. The deduplication read quantity refers to the unit counted in S32 based solely on the identification identifier. The system calculates the total number of valid RFID tags successfully read within a given time frame after deduplication, eliminating duplicate counts caused by repeated readings of the same tag and retaining only the number of individual fabric items. It also extracts the fabric count from the visual recognition dataset corresponding to the same single detection cycle, and simultaneously extracts the total number of RFID tags deduplicated by identity from the radio frequency dataset. This establishes a one-to-one correspondence between the two source quantity indicators and performs numerical comparison. This avoids the defects of inflated counts caused by directly using the original radio frequency reading count and the inability to verify visual targets and RFID tags one-to-one. Using the target quantity as the first verification threshold, it can quickly distinguish between two conditions: "consistent visual and radio frequency target quantities" and "deviations in the two source quantities," providing a basis for subsequent quantitative comparison.

[0051] Step S42 is used to bind the visual information of a single fabric to its RFID identification code based on spatial co-occurrence constraints after the visual and RFID dual-source target quantity matching is successful, realizing the entity association mapping of the two types of heterogeneous data. Among them, successful association matching is the preliminary judgment conclusion after the quantity comparison in S41, which means that the total number of fabric individuals visually identified in the current detection area is equal to the total number of independent tags effectively read by RFID, and has the data basis to establish a one-to-one binding relationship. Co-occurrence relationship means that visual image acquisition and RFID RFID excitation are both applied to the same detection station, and the fabric targets appearing in the spatial area in the same time period have a natural spatiotemporal binding relationship with their embedded RFID tags, without cross-regional cross-reading or background-irrelevant tag interference. The structured data entry is the visual standardized data unit corresponding to a single shot generated in S24, which carries the category, quantity, and appearance of a single fabric. Status and various confidence levels; the identity identifier is the unique RFID code of a single fabric obtained from S31 parsing, which is the core index for distinguishing individuals in the radio frequency dataset; under the premise that the number of fabrics and the number of deduplication reads are equal, it is determined that the visual and radio frequency data are successfully associated and matched. Based on the spatiotemporal constraints of synchronous acquisition and coexistence in the same workstation, a unique, bidirectional one-to-one correspondence mapping relationship is established between each independent fabric structured entry in the visual recognition data and the identity identifier of the corresponding single fabric in the radio frequency dataset; this can avoid the defects of disordered confusion and inaccurate pairing between visual targets and RFID tags in mixed scenarios of multiple fabrics. Based on the spatiotemporal co-occurrence characteristics of the unified detection area, entity binding is completed, and the data link between visual dimension attribute information and radio frequency dimension unique traceability code is opened up, providing a complete and accurate cross-source association mapping foundation for subsequent weighted comprehensive judgment by integrating multi-dimensional confidence indicators.

[0052] Step S43 is used to perform weighted comprehensive judgment by integrating multi-dimensional confidence quantification indicators based on the vision-RF one-to-one association mapping relationship established in S42, distinguishing three different levels of confidence in the detection conditions and outputting differentiated judgment benchmarks and results. Among them, category confidence, quantity confidence, and appearance status confidence are visual recognition quantification confidence indicators encapsulated in S24, which respectively characterize the reliability of fabric category, target count, and appearance defect recognition results. The readable confidence level and comprehensive perception score are RF quality evaluation indicators calculated by S33 and S34 based on the number of RF reads and field strength signals, which intuitively reflect the stability of RFID tag communication reading and signal strength level. The first preset threshold is the engineering-calibrated comprehensive perception score threshold, used to identify tags with completely reliable RF signal quality. The second preset threshold is the appearance status confidence threshold, used to identify extreme conditions where the credibility of visual defect recognition results is insufficient.

[0053] The hierarchical decision-making logic is divided into three scenarios: The first scenario: If the readability confidence level of the corresponding identity identifier is at the sure level and the comprehensive perception score is higher than the first preset threshold, it means that the RFID tag reading is stable, the signal is sufficient, and the traceability coding is highly reliable. At this time, the RFID unique identity identifier is used as the core benchmark for fabric counting and traceability. At the same time, the fabric category and appearance status output by the visual branch are used to synchronously correct the basic attributes of the tag and output a highly reliable first judgment result. The second scenario: If the readability confidence level is doubtful or missing, it means that the radio frequency signal is fluctuating and the risk of missing reading is high. The RFID code is not reliable enough. The radio frequency tag is abandoned as the judgment benchmark. Instead, the quantity and appearance of the fabric obtained by the visual unit are used as the core judgment basis, and the second judgment result based on the visual data is output. The third scenario: If the number of fabrics obtained by S41 comparison is inconsistent with the number of deduplication reads, or if the readability confidence is at the missing level and the appearance confidence is lower than the second preset threshold, it means that the visual and radio frequency data cannot be mutually verified, and the credibility of the visual defect recognition result is insufficient. It is determined that the dual-source data association matching has failed, and the device tag abnormality alarm is triggered simultaneously, indicating that there are faults such as missing fabrics, damaged tags, detached tags, and radio frequency interference. The third judgment result of the abnormality category is output. By linking multi-branch confidence indicators of vision with radio frequency comprehensive perception scores and graded confidence levels for hierarchical weighted judgment, different data sources are selected as judgment benchmarks for different scenarios. This can solve the defects of poor fault tolerance of single vision or single RFID recognition and distortion of judgment results in interference scenarios. In high-reliability scenarios, RFID is used to achieve accurate traceability, while in radio frequency failure scenarios, vision is used as a backup for identification. When there is a serious conflict between the two sources, the alarm is actively alarmed to troubleshoot the fault. It takes into account multiple needs such as fabric counting, attribute recognition, and tag fault monitoring, and completes the final comprehensive judgment output of the fabric status within a single detection cycle.

[0054] In one embodiment, step S40, which involves determining the abnormal attribution type of each medical fabric based on the judgment result, and using the identity identifier as an index to associate the current flow node and the abnormal attribution type with the corresponding identity identifier to construct a fabric state matrix, specifically includes: S441. If the judgment result is the first judgment result, mark the abnormal attribution type as normal identification and update the current circulation node to RFID inventory completion. S442. If the judgment result is the second judgment result, the quantity and appearance of the fabric are used as the judgment criteria, and the abnormal attribution type is marked as visual recognition; if the appearance is abnormal, the current flow node is updated to await manual re-inspection; if the appearance is normal, the current flow node is updated to visual review completed. S443. If the judgment result is the third judgment result, mark the abnormal attribution type as RFID tag abnormal and update the current circulation node to equipment abnormal. S45. Using the identity identifier as the index key, construct a status data entry containing the anomaly attribution type, the current flow node, and the timestamp, and associate the status data entry with the corresponding identity identifier; S46. Integrate all the status data entries corresponding to the identity identifiers to construct the fabric status matrix.

[0055] In this embodiment, as described in steps S441-S46 above, specifically: Steps S441-S443 are used to complete the marking of single fabric anomaly type and automatic update of business flow nodes based on the three types of differentiated comprehensive judgment results output by S43, so as to realize automatic diversion and control of fabric processing flow under different identification conditions. Among them, the anomaly attribution type is a classification mark used to distinguish the root cause of the anomaly in the current detection data, including three types of tags: normal identification, visual identification, and RFID tag anomaly. The current flow node is a business processing node set to match the entire process of medical fabric washing and counting, which is used to guide the subsequent automated processing or manual intervention process of the fabric, including four node statuses: RFID counting completed, awaiting manual re-inspection, visual verification completed, and equipment anomaly. The layered processing logic is as follows: In the first working condition, when S43 outputs the first judgment result, it means that the RFID tag's credibility level is at the certainty level, the comprehensive perception score meets the standard, the visual and RFID data matching degree is high, and there is no identification distortion problem. The abnormal attribution type of the corresponding fabric is marked as normal identification, and the business flow node of the fabric is updated to RFID counting completed. It can directly enter the next automated sorting process. In the second scenario, when S43 outputs the second judgment result, the RFID tag reading quality is insufficient and cannot be used as a traceability benchmark. The judgment is completed solely based on visual recognition data, and the abnormality attribution type is uniformly marked as visual recognition. On this basis, the appearance status branch is further subdivided: if the visual recognition outputs an abnormality type such as damage or stains, the flow node is updated to await manual re-inspection, and staff are arranged to check the fabric defects on-site; if the visual recognition determines that the appearance status is intact and without defects, the flow node is updated to visual verification completed, and compliance inventory is completed based on the visual data. In the third working condition, when S43 outputs the third judgment result, it means that the number of dual-source data is mismatched or the tag is invalid and the visual credibility is insufficient. The judgment data matching fails and an abnormal alarm is triggered. The abnormal attribution type is marked as RFID tag abnormality, the flow node is synchronously updated to equipment abnormality, the automated inventory process is suspended, and staff are prompted to check for problems such as tag falling off, equipment radio frequency failure, and missing fabric. By configuring differentiated anomaly attribution tags and flow nodes for the three-layer judgment results, the process can be automatically diverted to identify three scenarios: normal, visual fallback, and radio frequency fault. This avoids the defects of uniform flow for all working conditions and the inability to accurately screen and distinguish abnormal fabrics. It clarifies the source of each fabric anomaly and automatically assigns automated processing or manual intervention procedures. This provides a complete and hierarchical process tagging foundation for S45 to build standardized status data entries that are bound to identity identifiers, anomaly types, and flow nodes.

[0056] Step S45 is used to complete the structured binding and index archiving of the full-dimensional status information of a single fabric, constructing standardized status data entries that are traceable and time-series traceable, providing a single data unit for global matrix integration; among them, the identity identifier is a unique traceability code that binds a single medical fabric, serving as a unique index key for global data association, ensuring that each status data can correspond to a single independent fabric; the anomaly attribution type is the fault classification label output by S44, including three attribution results: normal identification, visual identification, and RFID tag anomaly, used to locate the source of the problem in a single detection; the current flow node is the business process node updated by S44, covering four flow states: RFID inventory completed, visual verification completed, awaiting manual re-inspection, and equipment anomaly, determining the subsequent processing flow of the fabric; the timestamp is the completion of a single detection cycle. The system establishes a standardized time-series record for determining the status, accurately marking the moment of status generation and enabling full-process time-series traceability. Using the unique identifier corresponding to each fabric as a locking index, it uniformly collects the fabric's anomaly attribution type, latest flow node status, and detection timestamp. Independent status data entries are generated by encapsulating these entries according to preset standardized fields and forcibly associating them with the corresponding identifier. This achieves a one-to-one mapping and archiving of four dimensions of information: fabric identity, anomaly type, flow status, and detection time. This avoids the defects of mixed status data from multiple fabrics, lack of independent indexes, and inability to trace individual items. It also solves the technical shortcomings of traditional detection results that only output judgment conclusions, lack process traceability, lack time-series records, and lack fault classification. This provides a standardized, searchable, and traceable single-item data foundation for subsequently integrating global data to construct a complete fabric status matrix.

[0057] Step S46 is used to summarize and integrate all individual fabric status data entries, constructing a multi-dimensional, batch-searchable standardized fabric status matrix to achieve unified global management of all fabric transfer and anomaly information. The status data entries are standardized units encapsulated using identifiers as indexes, as defined in S45. Each entry integrates a unique identifier, anomaly attribution type, current transfer node, and a complete timestamp. The fabric status matrix is ​​a two-dimensional structured data table. The matrix rows correspond to the unique identifiers of different individual fabrics, while the matrix columns are uniformly configured with fixed fields for identifier, anomaly attribution type, transfer node, and detection timestamp, forming... A globally unified state storage medium with well-organized rows and columns and standardized fields; it aggregates all valid identity-bound state data entries within a single inspection cycle, performs batch integration, sorting, and filling according to unified row and column rules, and generates a complete fabric state matrix; it avoids the shortcomings of scattered storage of multiple fabric state data, lack of a unified visualization medium, and low efficiency of batch retrieval, and aggregates discrete single-item state information into an integrated matrix data source, which is convenient for the upper-level system to batch read, statistically sort, export abnormal reports, and trace the entire process, providing globally unified, intuitive, and quantifiable data support for medical fabric inventory, defect re-inspection, RFID fault diagnosis, and flow process scheduling.

[0058] In one embodiment, a vision-based RFID identification and management system for medical fabrics is provided, which corresponds to the vision-based RFID identification and management method for medical fabrics described in the above embodiments. The vision-based RFID identification and management system for medical fabrics includes: The fabric positioning module is used to place the medical fabric to be processed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. The visual recognition module is used to identify the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model using the visual detection unit of the detection device. It outputs the category confidence, quantity confidence, and appearance confidence respectively, and generates a visual recognition dataset. The radio frequency acquisition module is used to read the RFID tags embedded in the TCM fabric in the preset detection area with radio frequency carrier signals within a single detection cycle using the RFID reading and writing unit of the detection device. It counts the number of successful reads per unit time and extracts the instantaneous field strength value of the reflected signal to generate a radio frequency dataset including unique department information and identity identifier. Based on the number of successful reads and the instantaneous field strength value, it evaluates the readability confidence level of each identity identifier. The data processing module is used to associate and match the visual recognition dataset and the radio frequency dataset, perform weighted judgment according to category confidence, quantity confidence, appearance status confidence and readability confidence level, determine the abnormal attribution type of each medical fabric based on the judgment result, and use the identity identifier as an index to associate the current flow node and abnormal attribution type with the corresponding identity identifier to construct the fabric status matrix. The periodic update module is used to update the fabric lifecycle records in the background management system based on the flow nodes.

[0059] For specific limitations regarding the vision-based RFID identification and management system for medical fabrics, please refer to the limitations of the vision-based RFID identification and management method for medical fabrics mentioned above, which will not be repeated here. Each module in the aforementioned vision-based RFID identification and management system for medical fabrics can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0060] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a vision-based RFID identification and management method for medical fabrics.

[0061] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a visual detection-based RFID identification and management method for medical fabrics.

[0062] As needed, in one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements a vision-based RFID identification and management method for medical fabrics.

[0063] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0064] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0065] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for RFID identification and management of medical fabrics based on visual inspection, characterized in that, Includes the following steps: The medical fabric to be processed is placed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. The visual detection unit of the detection device identifies the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model, and outputs the category confidence, quantity confidence, and appearance confidence respectively to generate a visual recognition dataset. The RFID reading and writing unit of the detection device reads the RFID tags embedded in the TCM fabric in the preset detection area with radio frequency carrier signals within a single detection cycle. The number of successful reads per unit time is counted and the instantaneous field strength value of the reflected signal is extracted to generate a radio frequency dataset including unique department information and identity identifier. The readability confidence level of each identity identifier is evaluated based on the number of successful reads and the instantaneous field strength value. The visual recognition dataset and the radio frequency dataset are associated and matched. Weighted judgment is performed according to category confidence, quantity confidence, appearance status confidence and readability confidence level. Based on the judgment results, the abnormal attribution type of each medical fabric is determined. Using the identity identifier as an index, the current circulation node and abnormal attribution type are associated with the corresponding identity identifier to construct the fabric status matrix. The fabric lifecycle record in the background management system is updated based on the aforementioned transfer node.

2. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 1, characterized in that, The construction of the pre-built multi-task branch fusion model specifically includes: A medical fabric sample set is constructed, which contains several images of medical fabrics in different appearance states, and each medical fabric image is labeled with a fabric category label, quantity bounding box, appearance state label and feature point label. A multi-branch model architecture is constructed, which includes a shared feature extraction layer and parallel-connected category recognition branch, quantity statistics branch, and appearance state analysis branch; A joint loss function is constructed, which is composed of a weighted sum of a cross-entropy loss term for constraining the category recognition branch, a smoothing absolute error loss term for constraining the quantity statistics branch, a focusing loss term for constraining the appearance state analysis branch, and a reprojection loss term for constraining the feature point labels. The medical fabric sample set is divided into a training set and a validation set. Based on the joint loss function, the multi-branch model architecture is iteratively trained using the training set and validated using the validation set. When the validation set loss value converges, the trained multi-task branch fusion model is obtained.

3. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 2, characterized in that, The step of using the visual detection unit of the detection device to identify the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model, and outputting category confidence, quantity confidence, and appearance confidence respectively to generate a visual recognition dataset, specifically includes: The visual inspection unit of the inspection device acquires images of medical fabrics in a preset inspection area; The medical fabric image is input into a multi-task branch fusion model. The medical fabric image is then subjected to feature extraction and fusion through a shared feature extraction layer to obtain fused features including category-aware features, location-aware features, and anomaly-aware features. The fused features are input into the category recognition branch, the quantity statistics branch, and the appearance state analysis branch, respectively. The category recognition branch performs classification processing on the category-aware features and outputs the fabric category and the corresponding category confidence score. The quantity statistics branch performs bounding box regression processing on the location-aware features and outputs the fabric quantity and the corresponding quantity confidence score. The appearance state analysis branch performs anomaly classification processing on the anomaly-aware features and outputs the appearance state and the corresponding appearance state confidence score. The fabric category, fabric quantity, appearance status, and their respective confidence scores are encapsulated into structured data entries to generate a visual recognition dataset.

4. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 1, characterized in that, The step of using the RFID read / write unit of the detection device to read the RFID tags embedded in the TCM fabric in the preset detection area, generating a radio frequency dataset including unique department information and identification, and evaluating the readability confidence level of each identification, specifically includes: The RFID reader / writer unit of the detection device transmits radio frequency carrier signals at a preset power within a preset detection area, and receives the reflected signals returned by each RFID tag based on the radio frequency carrier signals, and parses them to obtain tag data containing unique department information and identity identifiers; Within a single detection cycle, the number of successful reads of the RFID tags per unit time is counted according to the identity identifier group, and the instantaneous field strength value corresponding to each identity identifier in the reflected signal is extracted. The reading stability feature is calculated based on the number of successful reads per unit time, and the signal strength feature is calculated based on the instantaneous field strength value. The comprehensive perception score of the RFID tag is calculated based on the reading stability feature and the signal strength feature. Based on the comprehensive perception score, the readability confidence level of the corresponding identity identifier is divided into a certain level, a doubtful level, and a missing level, generating an RF dataset containing unique department information, identity identifier, and readability confidence level.

5. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 4, characterized in that, The steps of calculating the reading stability feature based on the number of successful reads per unit time, calculating the signal strength feature based on the instantaneous field strength value, and calculating the comprehensive perception score of the RFID tag based on the reading stability feature and the signal strength feature specifically include: The dispersion of the number of successful reads per unit time is statistically analyzed, and the read stability feature is calculated. The read stability feature is inversely proportional to the variance of the number of successful reads. Within the sliding time window of a single detection cycle, the instantaneous field strength value is filtered to remove outlier spikes caused by electromagnetic interference, resulting in a smoothed field strength sequence. The signal strength feature is calculated based on the smoothed field strength sequence. The signal strength feature is the difference between the mean of the field strength sequence and the current environmental noise baseline. The current environmental noise baseline is the mean of the radio frequency noise floor collected by the detection device when no tag passes by or the initial noise floor in the first frame detection period. The comprehensive perception score of the RFID tag is obtained by multiplying the reading stability feature by the signal strength feature.

6. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 4, characterized in that, The step of associating and matching the visual recognition dataset with the radio frequency dataset, and performing a weighted judgment based on category confidence, quantity confidence, appearance status confidence, and readability confidence level, specifically includes: Extract the number of fabrics from the visual recognition dataset and the number of deduplicated reads of the identity tags from the radio frequency dataset, and compare the number of fabrics with the number of deduplicated reads, wherein the number of deduplicated reads is the number of RFID tags after deduplication by identity tag in the number of successful reads per unit time. When the number of fabrics matches the number of duplicate reads, the association matching is determined to be successful. Based on the co-occurrence relationship in the preset detection area, a one-to-one correspondence between the fabrics and identity identifiers corresponding to each structured data entry in the visual recognition dataset is established. Based on the association and matching relationship, the confidence levels of the category, quantity, and appearance status of the associated fabrics, as well as the readability confidence level of the associated identity identifier, are obtained and weighted together with the comprehensive perception score for judgment, and the judgment result is output: When the readability confidence level is at the certainty level and the comprehensive perception score is greater than the first preset threshold, it is determined that the identity identifier is used as the fabric counting benchmark, and the category and appearance status are used as the attribute correction benchmark, and the first determination result is output. When the readability confidence level is doubtful or missing, the quantity and appearance of the fabric are used as the judgment criteria, and a second judgment result is output. When the number of fabrics is inconsistent with the number of deduplicated reads, or when the readability confidence level is missing and the appearance status confidence level is lower than the second preset threshold, the association matching is determined to be unsuccessful and a tag anomaly alarm is triggered, and a third determination result is output.

7. The RFID identification and management method for medical fabrics based on visual inspection as described in claim 6, characterized in that, The step of determining the abnormal attribution type of each medical fabric based on the judgment result, and using the identity identifier as an index to associate the current flow node and abnormal attribution type with the corresponding identity identifier to construct the fabric state matrix specifically includes: If the judgment result is the first judgment result, the abnormal attribution type will be marked as identification normal, and the current circulation node will be updated to RFID inventory completion; If the judgment result is the second judgment result, the quantity and appearance of the fabric are used as the judgment criteria, and the abnormal attribution type is marked as visual recognition; if the appearance is abnormal, the current flow node is updated to await manual re-inspection; if the appearance is normal, the current flow node is updated to complete visual verification. If the judgment result is the third judgment result, the anomaly attribution type will be marked as RFID tag anomaly, and the current circulation node will be updated to equipment anomaly; Using the identity identifier as the index key, construct a status data entry containing the anomaly attribution type, the current flow node, and the timestamp, and associate the status data entry with the corresponding identity identifier; By integrating all the status data entries corresponding to the identity identifiers, a fabric status matrix is ​​constructed.

8. A vision-based RFID identification and management system for medical fabrics, used to implement the steps of the vision-based RFID identification and management method for medical fabrics as described in any one of claims 1-7, characterized in that, include: The fabric positioning module is used to place the medical fabric to be processed in a single-layer flat position in the preset detection area of ​​the detection device. The preset detection area includes a visual acquisition area and an RFID reading and writing area that overlap spatially. The visual recognition module is used to identify the category, quantity, and appearance of TCM fabrics in a preset detection area based on a pre-built multi-task branch fusion model using the visual detection unit of the detection device. It outputs the category confidence, quantity confidence, and appearance confidence respectively, and generates a visual recognition dataset. The radio frequency acquisition module is used to read the RFID tags embedded in the TCM fabric in the preset detection area with radio frequency carrier signals within a single detection cycle using the RFID reading and writing unit of the detection device. It counts the number of successful reads per unit time and extracts the instantaneous field strength value of the reflected signal to generate a radio frequency dataset including unique department information and identity identifier. Based on the number of successful reads and the instantaneous field strength value, it evaluates the readability confidence level of each identity identifier. The data processing module is used to associate and match the visual recognition dataset and the radio frequency dataset, perform weighted judgment according to category confidence, quantity confidence, appearance status confidence and readability confidence level, determine the abnormal attribution type of each medical fabric based on the judgment result, and use the identity identifier as an index to associate the current flow node and abnormal attribution type with the corresponding identity identifier to construct the fabric status matrix. The periodic update module is used to update the fabric lifecycle records in the background management system based on the flow nodes.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the RFID identification and management method for medical fabrics based on visual detection as described in any one of claims 1-7.

Images