Medical instrument storage scene identification method and system based on improved YOLOv11

Abstract

The invention discloses a medical instrument storage scene recognition method and system based on improved YOLOv11, and the system comprises a data collection module, a target recognition module, a scene analysis module, an alarm module and a data storage module, and the detection module is used for performing feature extraction on the received real-time image data through the C3k2 sub-module, the SPPF sub-module and the C2PSA sub-module in the backbone network to obtain feature information, inputting the feature information into a feature pyramid for feature fusion and enhancement, inputting the fused and enhanced features into the detection head to output a plurality of recognition results, and outputting the recognition results. According to the invention, accurate identification of various scenes in the medical instrument storage environment can be realized without manually adding environment road signs, and the efficiency and accuracy of storage management are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection, and particularly relates to a medical instrument warehouse scene recognition method and system based on improved YOLOv11. BACKGROUND

[0002] The warehouse management of medical instruments has extremely high requirements for safety and accuracy, and the traditional manual management mode is inefficient and prone to errors, which is difficult to meet the needs of modern warehouse. For example, there are various types of medical instruments with different shapes, and manual inventory is easy to miss or misjudge, and the inventory status cannot be monitored in real time. In the warehouse management of medical instruments, accurately identifying various scenes in the warehouse environment is of great significance to improve the efficiency of the warehouse, ensure the safe storage of medical instruments, and quickly in and out of the warehouse. With the development of computer vision and deep learning technology, it is possible to use advanced target detection models to realize the automatic identification of the warehouse environment of medical instruments. At present, the warehouse management of medical instruments mainly relies on manual inventory and simple monitoring systems. Manual inventory requires workers to enter the warehouse environment regularly to count and check medical instruments one by one, which is not only inefficient, but also prone to errors due to fatigue or negligence. Although the simple monitoring system can monitor the warehouse environment in real time, it cannot automatically identify the types, quantities and states of medical instruments, and cannot intelligently analyze and alarm abnormal situations in the warehouse environment. In addition, some traditional target detection models can achieve certain automatic identification functions, but they have low recognition accuracy, poor real-time performance, and insufficient ability to handle complex backgrounds and occlusion problems in the warehouse environment of medical instruments. The existing technology has the following shortcomings: 1. Manual inventory is inefficient and prone to errors, and cannot monitor the inventory status in real time; 2. The simple monitoring system cannot automatically identify the types, quantities and states of medical instruments, and cannot intelligently analyze and alarm abnormal situations; 3. The traditional target detection model has low recognition accuracy, poor real-time performance, and insufficient ability to handle complex backgrounds and occlusion problems in the warehouse environment of medical instruments. SUMMARY

[0003] (I) Invention purpose

[0004] The purpose of the present application is to provide a medical instrument warehouse scene recognition method and system based on improved YOLOv11, which can accurately identify various scenes in the warehouse environment of medical instruments without the need for human to add environmental landmarks, thereby improving the efficiency and accuracy of warehouse management.

[0005] (II) Technical solution

[0006] To solve the above problems, the first aspect of the present application provides a medical instrument warehouse scene recognition system based on improved YOLOv11, characterized in that it comprises:

[0007] The data acquisition module is configured to acquire real-time image data of the medical instrument storage scene and send the preprocessed real-time image data to the target recognition module.

[0008] The target recognition module is an improved YOLOv11 model trained, configured to perform feature extraction on the received real-time image data through a C3k2 submodule, an SPPF submodule and a C2PSA submodule in a backbone network to obtain feature information, input the feature information into a feature pyramid for feature fusion and enhancement, input the fused and enhanced feature into a detection head to output a plurality of recognition results, and transmit the recognition results to the scene analysis module.

[0009] The scene analysis module is configured to receive the recognition results and compare the recognition results with a plurality of preset scenes to determine whether the recognition results are abnormal.

[0010] The alarm module is configured to determine whether to trigger an alarm signal according to an abnormal condition of the recognition results and match a corresponding alarm signal according to an abnormal type, and generate an alarm record.

[0011] The data storage module is in communication connection with the target recognition module and the alarm module, and is configured to store the recognition results and the alarm record.

[0012] Further, the plurality of recognition results include a bounding box detection result, a class prediction result and a confidence prediction result.

[0013] Further, the C3k2 submodule uses a 2x2 convolution kernel to process the real-time image, and uses a cross-stage partial connection method, and the C2PSA module uses a parallel spatial attention mechanism.

[0014] Further, the improved YOLOv11 model is trained using COCO pre-training weights and updated and optimized through a stage-by-stage training strategy.

[0015] Further, in the training process, a BCE loss function and a CIOU loss function are constructed, and an Adam or SGD optimizer is used for optimization training.

[0016] Further, the bounding box detection result includes a plurality of bounding boxes predicted in each grid of each prediction layer, and each bounding box defines parameters including x and y coordinates of a center node, width and height, and the recognition result includes a plurality of prediction layers.

[0017] Further, the category prediction result comprises predicting the category of each bounding box, and the confidence prediction result comprises predicting the confidence of the target in each bounding box.

[0018] Further, the comparing the recognition result with the plurality of preset scenes to determine whether the recognition result is abnormal comprises:

[0019] extracting the bounding box detection result, the category prediction result and the confidence prediction result;

[0020] loading the preset scenes, and comparing the bounding box detection result, the category prediction result and the confidence prediction result with the preset scenes one by one;

[0021] if the matching is successful, it is determined that the recognition result is not abnormal, and if the matching fails, it is determined that the recognition result is abnormal.

[0022] Further, the system further comprises an environment sensor module, which is electrically connected with the target recognition module and the data storage module respectively.

[0023] In addition, the second aspect of the present application provides a medical instrument storage scene recognition method, which is realized based on the improved YOLOv11 medical instrument storage scene recognition system described in any of the above descriptions, and characterized in that the method comprises:

[0024] acquiring real-time image data of the medical instrument storage scene through a data acquisition module, and sending the preprocessed real-time image data to the target recognition module;

[0025] extracting features from the received real-time image data through the target recognition module to obtain feature information, inputting the feature information into a feature pyramid for feature fusion and enhancement, inputting the fused and enhanced features into a detection head to output a plurality of recognition results, and transmitting the recognition results to the scene analysis module, the target recognition module being an improved YOLOv11 model after training, the improved YOLOv11 model comprising a backbone network, a feature pyramid and a detection head; the backbone network comprising a C3k2 submodule, an SPPF submodule and a C2PSA submodule;

[0026] receiving the recognition result through the scene analysis module, and comparing the recognition result with a plurality of preset scenes to determine whether the recognition result is abnormal;

[0027] determining whether to trigger an alarm signal according to the abnormal condition of the recognition result and matching the corresponding alarm signal according to the abnormal type through the alarm module, and generating an alarm record;

[0028] The data storage module is in communication connection with the target identification module and the alarm module respectively to store the identification result and alarm record.

[0029] (III) Beneficial Effects

[0030] The above technical scheme of the present application has the following beneficial technical effects: the present application provides a medical instrument storage scene identification method and system based on improved YOLOv11, which does not need to rely on manually set environment identification or markers, reduces management cost, and improves the overall cleanliness and aesthetics of the storage environment. The system comprises a data acquisition module, a target identification module, a scene analysis module, an alarm module and a data storage module. The target identification module is an improved YOLOv11 model after training, which is used to extract features from the received real-time image data through the C3k2 submodule, SPPF submodule and C2PSA submodule in the backbone network to obtain feature information, then input the feature information into the feature pyramid for feature fusion and enhancement, input the fused and enhanced features into the detection head to output multiple identification results, and transmit the identification results to the scene analysis module. The present application can process image data in the storage environment in real time and quickly output identification results to provide timely and accurate information support for storage management. By introducing the YOLOv11 model and its improved components (C3k2, SPPF, C2PSA), the present application can realize high-precision identification of medical instruments and their storage scenes, significantly improving the efficiency and accuracy of storage management. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a medical instrument storage scene identification system based on improved YOLOv11 of the present application;

[0032] Figure 2 is a key architecture module diagram of the improved YOLOv11 of the specific embodiment of the present application;

[0033] Figure 3 is a medical instrument storage scene identification system flow chart based on improved YOLOv11 of the specific embodiment of the present application;

[0034] Figure 4 is an alarm module diagram of the specific embodiment of the present application;

[0035] Figure 5 is a data storage module diagram of the specific embodiment of the present application. DETAILED DESCRIPTION

[0036] To make the objectives, technical solutions, and advantages of the present application clearer, further detailed descriptions of the present application will be given below with reference to the embodiments and in conjunction with the drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present application. Moreover, in the following description, descriptions of well-known structures and techniques are omitted to avoid unnecessary confusion of the concepts of the present application.

[0037] As Figure 1 shown, the first aspect of the present application provides an improved YOLOv11-based medical instrument storage scene recognition system, comprising:

[0038] A data acquisition module 1 is configured to acquire real-time image data of the medical instrument storage scene and send the preprocessed real-time image data to the target recognition module. The data acquisition module 1 is configured to acquire image data in the medical instrument storage environment in real time, including a plurality of cameras deployed at key positions in the storage environment for multi-angle image data acquisition. These cameras are taken from different angles to ensure comprehensive coverage of the storage environment. The collected image data is first preprocessed, including adjusting the image size to adapt to the input requirements of the YOLOv11 model (e.g., 640x640 pixels), and normalization and other operations to improve the recognition performance of the model.

[0039] The target recognition module 2 is a trained improved YOLOv11 model. The improved YOLOv11 model is trained to recognize medical instruments and their states in the image. The preprocessed image data is input into the trained improved YOLOv11 model to output the recognition result. As Figure 2As shown, the target recognition module 2 specifically extracts features from the received real-time image data through the C3k2 submodule, SPPF submodule, and C2PSA submodule in the backbone network to obtain feature information. The model extracts and analyzes features of the image through its improved architecture (including components such as C3k2, SPPF, and C2PSA), inputs the feature information into the feature pyramid for feature fusion and enhancement, inputs the fused and enhanced features into the detection head to output multiple recognition results, and transmits the recognition results to the scene analysis module. The improved YOLOv11 model includes a backbone network, a feature pyramid, and a detection head. In the target recognition module 2, the C3k2 (Cross Stage Partial with kernel size 2), SPPF (SpatialPyramid Pooling Fast), and C2PSA (Convolutional block with Parallel Spatial Attention) components are introduced to significantly improve the performance of the model. C3k2 uses cross-stage partial connection and a small convolution kernel (kernel size 2) to enhance feature extraction capability, reduce computational complexity and model parameters, and improve running efficiency. SPPF effectively improves the model's detection ability for targets of different scales through multi-scale feature fusion, especially when dealing with small targets and complex backgrounds, thereby improving the accuracy and robustness of detection. The C2PSA component uses a parallel spatial attention mechanism to enable the model to automatically identify and focus on key feature regions in the image, such as the outlines and labels of medical devices, further enhancing feature extraction efficiency and accuracy. These improvements not only improve the recognition accuracy and robustness of the model but also optimize computational efficiency, enabling the system to better adapt to complex scenarios in the medical device storage environment and achieve high-precision, real-time, and robust scene recognition, significantly improving the efficiency and safety of medical device storage management.

[0040] The multiple recognition results include bounding box detection results, class prediction results, and confidence prediction results. The bounding box detection results include predicting multiple bounding boxes in each grid of each prediction layer, and the parameters defined by each bounding box include the x and y coordinates of the center node, the width, and the height. The recognition results include multiple prediction layers. The improved YOLOv11 model outputs the bounding box coordinates (x, y, width, height) of each detected target, which are equivalent to the x and y coordinates of the center node, the width, and the height. These coordinates define the position and size of the target in the image. The improved YOLOv11 model predicts multiple bounding boxes in each grid (or cell) of each prediction layer. Each bounding box is defined by four parameters: the x and y coordinates of the center node, the width, and the height. These parameters are predicted through the regression branch of the model, and are optimized using a loss function (such as CIOU loss) to ensure that the bounding box can accurately locate the target.

[0041] The class prediction results include predicting the class of each bounding box, and the model simultaneously predicts the class to which each bounding box belongs, i.e., the type of medical device (such as valve, bearing, gear, etc.). Class prediction is achieved through the classification branch of the model, and the output is usually converted into a probability distribution using a softmax function to determine the most likely class of each bounding box. The accuracy of class prediction depends on high-quality labeled data and effective training strategies.

[0042] The confidence prediction results include detecting the confidence of the target within each bounding box. The model also outputs the confidence of each bounding box, which represents the degree of confidence of the model in detecting the target within the bounding box. Confidence is a value between 0 and 1, and the higher the value, the more confident the model is in the detection result. Confidence prediction is achieved through the confidence branch of the model, and the output is usually converted into a value between 0 and 1 using a sigmoid function. The calculation of confidence takes into account the accuracy of the bounding box and the accuracy of the class prediction. During the training process, the accuracy of confidence prediction is improved by optimizing the loss function (such as BCE loss).

[0043] The improved YOLOv11 model is trained using COCO pre-trained weights, and is updated and optimized through a phased training strategy. During the training process, the BCE loss function and the CIOU loss function are constructed, and the Adam or SGD optimizer is used for optimization training. The COCO pre-trained weights are used as the basis for fine-tuning the model to focus on the features of medical devices and warehouse environments. The YOLOv11 pre-trained weight file (such as `yolov11_coco.pt`) on the COCO2017 dataset is downloaded from the official channel (such as GitHub repository). The specific optimization strategy using the COCO dataset pre-trained weights is as follows:

[0044] 1) Fine-tuning strategy for transfer learning: Based on the pre-trained model, adjust the network parameters with a small amount of medical device dataset, so as to focus on the characteristics of medical devices and warehouse environment. Transfer learning is an effective strategy that can utilize the general features learned by pre-trained models on large-scale datasets (such as COCO) to adapt to specific tasks (such as medical device identification) through fine-tuning. The fine-tuning steps include: A. Load pre-trained weights: Load the YOLOv11 weights pre-trained on the COCO dataset; B. Data annotation: Assuming that the medical device dataset has been collected and labeled, the annotation content includes bounding box and class label.

[0045] 2) Training strategy: Use a small learning rate for fine-tuning, first freeze most layers of the pre-trained model, only train the last few layers, then gradually unfreeze more layers for training, as follows:

[0046] A, First, take high-quality image data of medical devices and other items in the warehouse environment, including medical device images under different angles, lighting, and occlusion conditions. Label the bounding box and class label of each component, use annotation tools (such as LabelImg, CVAT, Roboflow) to annotate the image, the annotation content includes bounding box (Bounding Box) and class label (such as valve, bearing, gear, etc.). For images related to environmental monitoring, you also need to label the location and type of abnormal environmental conditions (such as high-temperature areas, humidity abnormal areas), and use data augmentation techniques (such as simulating low light, occlusion, and multi-angle shooting) to improve the generalization ability of the model. Divide the dataset into training, validation, and testing groups.

[0047] B, Use the annotated image data to train the improved YOLOv11 model, adjust the learning rate, optimizer and other hyperparameters to make the model better learn the characteristics of medical devices and warehouse scenarios, each step is as follows:

[0048] B1, Learning rate (Learning Rate) adjustment: Use a small learning rate (such as 1e4) for fine-tuning to avoid too large an impact on pre-trained weights. Learning rate determines the step size of parameter updates in the training process. A smaller learning rate can ensure that the model does not skip the optimal solution during fine-tuning due to a large step size, so that it can more stably converge to better parameter values.

[0049] B2: Staged Training: In the early stage of training, most layers of the pre-trained model are frozen, and only the last few layers (such as the detection head) are trained. As training progresses, more layers are gradually unfrozen for training. Freezing most layers of the pre-trained model can preserve the general features learned by the pre-trained model on the COCO dataset, which are very important for the target detection task. Gradually unfreezing more layers can ensure that the model gradually adapts to the new task during fine-tuning, avoiding large-scale adjustment of the entire model at the beginning, thereby improving the stability and convergence speed of the model.

[0050] B3: Optimizer Selection: Use Adam or SGD optimizer combined with learning rate scheduler (such as StepLR, CosineAnnealingLR) for optimization. Adam optimizer: Adam optimizer combines the advantages of momentum and adaptive learning rate, which can converge faster and is relatively insensitive to the selection of hyperparameters. SGD optimizer: SGD optimizer can provide more stable convergence in the later stage of training, especially in the fine-tuning stage, combined with learning rate scheduler to better adjust the learning rate to avoid premature convergence to local optimal solution. Learning rate scheduler (such as StepLR, CosineAnnealingLR) can dynamically adjust the learning rate according to the training progress, for example, using a larger learning rate in the early stage of training to quickly converge, and then gradually reducing the learning rate in the later stage of training to improve the accuracy of the model.

[0051] B4: Loss Function Adjustment: Use classification BCE loss (Binary Cross-Entropy Loss) and regression CIOU loss (Complete Intersection over Union Loss) to improve the classification and positioning accuracy of the model. Classification BCE loss: used for classification tasks, calculates the binary cross-entropy loss between predicted categories and true categories, ensuring that the model can accurately classify medical devices.

[0052] Regression CIOU loss: used for bounding box regression tasks, calculates the complete intersection over union loss between predicted bounding boxes and true bounding boxes, not only considering the overlapping area of bounding boxes, but also considering the difference in shape and size, thereby improving the positioning accuracy of bounding boxes.

[0053] B5: Data Augmentation Adjustment Method: Generate more training samples through data augmentation techniques such as random cropping, flipping, rotation, color adjustment, etc. Data augmentation can increase the generalization ability of the model, making it better adapt to different lighting conditions, occlusion situations, background interference and other complex scenarios, thereby improving the robustness of the model in actual environment.

[0054] The scene analysis module 3, as shown in Figure 3 receives the identification results and compares them with a number of preset scenes to determine whether the identification results are abnormal. The identification results (bounding box detection results, class prediction results, and confidence prediction results) are passed to the scene analysis module, which compares and analyzes the identification results with the preset scene rules (preset scenes) to determine whether the scene is abnormal. The comparison of the identification results with a number of preset scenes to determine whether the identification results are abnormal includes the following steps:

[0055] (1) Extract the bounding box detection results, class prediction results, and confidence prediction results. Extract the identification results: extract the bounding box coordinates, class, and confidence of each detected target from the output of the YOLOv11 model.

[0056] (2) Load the preset scenes and compare the bounding box detection results, class prediction results, and confidence prediction results with the preset scenes one by one. Load the system's preset scene rules, which define the standards for normal and abnormal scenes. Compare each identification result with the preset rules one by one to determine whether it meets the standards for normal scenes.

[0057] (3) If the match is successful, determine that the identification results are not abnormal; if the match fails, determine that the identification results are abnormal. As shown in Figure 4 , the abnormal situations include inventory abnormalities, environmental abnormalities, and personnel abnormalities, and the determination standards for each abnormal situation are as follows:

[0058] ① Inventory anomaly classification and determination standards:

[0059] Quantity discrepancy: the number of medical devices identified does not match the inventory record.

[0060] Position anomaly: the placement position of the medical device exceeds the preset area.

[0061] Class error: the class of the medical device identified does not match the inventory record.

[0062] Example: the system identifies that the number of medical devices on a shelf is less than the inventory record, triggering an inventory quantity anomaly alarm.

[0063] The system identifies that a medical device is placed on the wrong shelf, triggering a location anomaly alarm.

[0064] ② Environment anomaly classification and determination criteria:

[0065] Temperature anomaly: Temperature sensor data exceeds a set threshold (e.g., 30°C).

[0066] Humidity anomaly: Humidity sensor data is below or above a set threshold (e.g., humidity is below 40% or above 70%).

[0067] Example: The system detects that the temperature of the warehouse environment reaches 32°C, triggering a temperature anomaly alarm.

[0068] The system detects that the humidity of the warehouse environment is 75%, triggering a humidity anomaly alarm.

[0069] ③ Personnel anomaly classification and determination criteria:

[0070] Illegal intrusion: Detection of unauthorized personnel entering restricted areas.

[0071] Improper operation: Detection of personnel's operation in the warehouse environment not in accordance with regulations (e.g., not wearing protective equipment).

[0072] Example: The system detects that an unauthorized person enters a restricted area, triggering an illegal intrusion alarm.

[0073] The system detects that a staff member does not wear protective equipment when operating a medical device, triggering an improper operation alarm.

[0074] Alarm module 4, for determining whether to trigger an alarm signal according to the abnormal condition of the identification result, and generating an alarm record according to the corresponding alarm signal matched according to the type of anomaly. If the scene analysis module detects an abnormal situation, such as the placement position of a medical device exceeding the preset area or the environmental parameter being abnormal, the alarm module will be triggered to issue a visual, sound, or message push alarm. According to the output result of the scene analysis module, if the identification result does not match the preset rules, it is determined as an abnormal scene, and the scene analysis module will output an abnormal signal. If an abnormal situation is detected, an alarm signal is triggered, and after the alarm module receives the abnormal signal, it triggers the corresponding alarm function according to the type of anomaly (inventory anomaly, environmental anomaly, personnel anomaly, etc.), such as Figure 4 As shown, it has visual alarm, sound alarm, and message push functions, and can timely remind the management personnel to handle the abnormal situation. The alarm module has a built-in message push function, which can push the alarm information to the management personnel through SMS, email, or instant messaging tools (such as WeChat, DingTalk, etc.). The push content includes alarm type, alarm time, specific description, and other detailed information, ensuring that the management personnel can quickly understand the abnormal situation and take appropriate measures. The transmission of each alarm signal is as follows:

[0075] 1) Visual alarm: Use monitoring system interface (such as computer screen, monitoring display screen), alarm indicator light and other devices, connection mode: the monitoring system interface is connected to the server through the network, and the monitoring picture and alarm information are displayed in real time, the alarm indicator light is connected to the alarm module through the GPIO interface, and the alarm module controls the on-off of the indicator light through the GPIO interface. When the alarm module receives an abnormal signal, it controls the alarm indicator light to turn on through the GPIO interface, and displays a red alarm sign and an abnormal information prompt box on the monitoring system interface. The server runs the YOLOv11 recognition module and the scene analysis module, processes image data and performs abnormal analysis. In the internal structure of the alarm module, the input interface receives the abnormal signal of the scene analysis module, and the control unit triggers the corresponding alarm function according to the abnormal type. The output interface is a GPIO interface for controlling the alarm indicator light, and an audio interface for controlling the loudspeaker. The network interface is connected to the message pushing module to realize information pushing.

[0076] 2) Sound alarm: Use devices such as loudspeakers, which are connected to the alarm module through the audio interface. The alarm module controls the loudspeaker to emit a preset alarm sound through the audio interface when it receives an abnormal signal, reminding the manager to pay attention.

[0077] The specific process of the alarm is as follows: real-time image data is collected by the monitoring camera deployed at the key position of the warehouse environment, and the image data is transmitted to the server through wired or wireless network. The server runs the YOLOv11 recognition module and the scene analysis module, processes the image data and performs abnormal analysis. The alarm module receives the abnormal signal of the scene analysis module and triggers the alarm function. The monitoring system interface is connected to the server through the network, and the monitoring picture and alarm information are displayed in real time. The alarm indicator light is connected to the alarm module through the GPIO interface, receives the control signal and turns on. The loudspeaker is connected to the alarm module through the audio interface, receives the control signal and emits the alarm sound. The message pushing module is connected to the alarm module through the network, and pushes the alarm information to the manager.

[0078] The data storage module 5 is in communication connection with the target recognition module 2 and the alarm module 4 respectively, and is used for storing the recognition results and alarm records. All image data, recognition results and alarm records are stored in the data storage module (database) for subsequent query and analysis. The image data of the medical instrument, the recognition results and the alarm records are also stored, and the environmental sensor data (such as temperature, humidity) are also stored and associated with the image data for environmental monitoring function.

[0079] The data storage module 5 includes functions such as storage, associative storage, and data query, among which real-time storage: the server stores the collected image data, recognition results, and alarm records into the database in real time; associative storage: the image data is stored in association with the recognition results, alarm records, and environmental data, facilitating subsequent query and analysis, such as shown in Figure 5 Data query: the database provides flexible data query functions, and the administrator can query historical data according to time range, alarm type, and other conditions to quickly obtain the required information. In addition, the connection relationships of the database with other devices are as follows:

[0080] A, server and database:

[0081] Network connection: the server is connected to the database server through a local area network (LAN) or a wide area network (WAN), and usually uses TCP / IP protocol for communication.

[0082] Database management system (DBMS): the database management system (such as MySQL, PostgreSQL, MongoDB, etc.) running on the server interacts with the database server through a network interface.

[0083] B, monitoring camera and server:

[0084] Network connection: the monitoring camera is connected to the server through wired or wireless network, and the collected image data is transmitted to the server in real time.

[0085] C, alarm module and server:

[0086] GPIO interface: the alarm module is connected to the server through the GPIO interface, and the server controls the visual and sound alarm functions of the alarm module through the GPIO interface.

[0087] Network interface: the alarm module is connected to the server through the network interface for message push function.

[0088] The system also includes an environmental sensor module 5, which is electrically connected to the target recognition module 2 and the data storage module 5. The environmental sensor module (such as temperature sensor, humidity sensor) is connected to the server through wired or wireless network, and the collected environmental data is transmitted to the server in real time.

[0089] On the other hand, the present application provides a medical instrument storage scene recognition method, which is based on any one of the above-mentioned medical instrument storage scene recognition systems based on improved YOLOv11. The method includes:

[0090] S1, collecting real-time image data of the medical instrument storage scene through a data acquisition module, and sending the preprocessed real-time image data to the target recognition module;

[0091] S2, extracting features from the received real-time image data through the target recognition module to obtain feature information, and inputting the feature information into a feature pyramid for feature fusion and enhancement, and inputting the fused and enhanced features into a detection head to output a plurality of recognition results, and transmitting the recognition results to the scene analysis module, the target recognition module being a trained improved YOLOv11 model, the improved YOLOv11 model including a backbone network, a feature pyramid and a detection head; the backbone network includes a C3k2 submodule, an SPPF submodule and a C2PSA submodule;

[0092] S3, receiving the recognition results through the scene analysis module, and comparing the recognition results with a plurality of preset scenes to determine whether the recognition results are abnormal;

[0093] S4, determining whether to trigger an alarm signal according to the abnormal condition of the recognition result and matching the corresponding alarm signal according to the abnormal type through the alarm module, and generating an alarm record;

[0094] S5, the data storage module is respectively connected with the target recognition module and the alarm module in communication to store the recognition results and the alarm record.

[0095] The application provides a medical instrument storage scene recognition method and system based on improved YOLOv11, which has the following advantages:

[0096] 1. High precision recognition: based on the deep improvement and optimization of YOLOv11 model, high precision recognition of medical instruments and their storage scenes can be realized, which significantly improves the efficiency and accuracy of storage management. The introduction of YOLOv11 model and its improved components (C3k2, SPPF, C2PSA) improves the precision and real-time performance of medical instrument storage environment scene recognition.

[0097] 2. Strong robustness and adaptability: through the carefully designed data preprocessing process and high-quality labeled data set, the complex and changeable storage environment can be effectively dealt with, and excellent robustness is shown.

[0098] 3. Real-time and fast response: capable of processing image data in the warehouse environment in real time and quickly outputting recognition results to provide timely and accurate information support for warehouse management. Multi-angle camera deployment and image acquisition, combined with environmental sensor data, achieve comprehensive warehouse environment monitoring. The alarm module integrates visual alarm, sound alarm and message push functions to timely remind managers to handle abnormal situations. Efficient storage and management of image data, recognition results and alarm records are achieved.

[0099] 4. No need for manual identification: no need to rely on manually set environment identifiers or markers, reducing management costs and improving the overall cleanliness and aesthetics of the warehouse environment.

[0100] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by programs instructing relevant hardware, and the programs can be stored in a computer-readable storage medium. When the program is executed, it includes the flow of the above-mentioned embodiments of each method. Among them, the storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc. The steps in the embodiment methods of the present application can be adjusted, combined and reduced in sequence according to actual needs. The modules in the embodiment system of the present application can be combined, divided and reduced according to actual needs.

Claims

1. A medical device warehousing scene recognition system based on an improved YOLOv11, characterized in that, include: The data acquisition module is used to acquire real-time image data of the medical device storage scene, and send the pre-processed real-time image data to the target recognition module; The target recognition module is a trained improved YOLOv11 model, which is used to extract features from the received real-time image data through the C3k2 submodule, SPPF submodule and C2PSA submodule in the backbone network to obtain feature information. The feature information is then input into the feature pyramid for feature fusion and enhancement. The fused and enhanced features are then input into the detection head to output multiple recognition results. The recognition results are then transmitted to the scene analysis module. The improved YOLOv11 model includes a backbone network, a feature pyramid and a detection head. The scene analysis module is used to receive the recognition result and compare the recognition result with several preset scenes to determine whether the recognition result is abnormal. The alarm module is used to determine whether to trigger an alarm signal based on the abnormality of the identification result, match the corresponding alarm signal according to the abnormality type, and generate an alarm record. The data storage module is communicatively connected to the target recognition module and the alarm module, respectively, and is used to store the recognition results and alarm records.

2. The medical device warehousing scene recognition system based on improved YOLOv11 according to claim 1, characterized in that, The recognition results include bounding box detection results, category prediction results, and confidence prediction results.

3. The medical device warehousing scene recognition system based on improved YOLOv11 according to claim 1, characterized in that, The C3k2 submodule uses a 2×2 convolution kernel to process the real-time image and adopts a cross-stage partial connection method. The C2PSA module adopts a parallel spatial attention mechanism.

4. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 1, characterized in that, The improved YOLOv11 model was trained using COCO pre-trained weights and then updated and optimized using a phased training strategy.

5. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 4, characterized in that, During the training process, the BCE loss function and CIOU loss function are constructed, and the Adam or SGD optimizer is used for optimization training.

6. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 2, characterized in that, The bounding box detection results include: predicting multiple bounding boxes for each grid in each prediction layer, and defining parameters for each bounding box including: the x and y coordinates of the center node, width, and height. The recognition results include multiple prediction layers.

7. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 2, characterized in that, The category prediction result includes predicting the category of each bounding box, and the confidence prediction result includes performing confidence detection on the targets within each bounding box.

8. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 1, characterized in that, The step of comparing the recognition result with several preset scenarios to determine whether the recognition result is abnormal includes: Extract the bounding box detection results, category prediction results, and confidence prediction results; Load a preset scene, and compare the bounding box detection results, category prediction results, and confidence prediction results with the preset scene one by one; If the match is successful, the recognition result is considered to be normal; if the match fails, the recognition result is considered to be abnormal.

9. The medical device warehouse scene recognition system based on improved YOLOv11 according to claim 1, characterized in that, The system also includes an environmental sensor module, which is electrically connected to the target recognition module and the data storage module, respectively.

10. A method for identifying medical device storage scenarios, said method being implemented based on the medical device storage scenario identification system based on the improved YOLOv11 as described in any one of claims 1-9, characterized in that, The method includes: The data acquisition module collects real-time image data of the medical device storage scene, and sends the pre-processed real-time image data to the target recognition module. The target recognition module extracts features from the received real-time image data to obtain feature information. This feature information is then input into a feature pyramid for feature fusion and enhancement. The fused and enhanced features are then input into a detection head to output multiple recognition results, which are then transmitted to the scene analysis module. The target recognition module is a trained improved YOLOv11 model, which includes a backbone network, a feature pyramid, and a detection head. The backbone network includes a C3k2 submodule, an SPPF submodule, and a C2PSA submodule. The scene analysis module receives the recognition results and compares them with several preset scenes to determine whether the recognition results are abnormal. The alarm module determines whether to trigger an alarm signal based on the abnormality of the identification results, matches the corresponding alarm signal according to the abnormality type, and generates an alarm record. The data storage module is connected to the target recognition module and the alarm module respectively to store the recognition results and alarm records.

Images