Field medical sharp instrument recycling management method and system based on module anti-permeation structure

Abstract

The present invention provides a method and system for recycling and management of field medical sharps based on a module anti-permeation structure, relating to the technical field of image analysis, including: identifying field medical sharps through multi-modal image fusion and a deep learning model, extracting key features for precise positioning, segmentation and semantic analysis, and classifying them by combining with a causal inference network to obtain category and risk level information. Based on the classification information and risk level information, an improved spatio-temporal graph neural network and a hierarchical reinforcement learning algorithm are used, combined with an improved auction mechanism and a security constraint model, to generate and optimize a recycling scheduling strategy. Through an improved ant colony algorithm and a heuristic search strategy, combined with a Pareto optimization algorithm, an optimal recycling route is planned.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and particularly to a method and system for recycling and management of field medical sharps based on a modular anti-permeation structure. Background Art

[0002] The safe recycling and management of field medical sharps is an important link in the field of medical and health. If medical sharps such as syringes, needles, and surgical blades are not properly handled, they will pose a serious threat to the environment and public health, and may lead to accidental injuries and disease transmission. Therefore, how to effectively identify, classify, and recycle field medical sharps is crucial for ensuring public health safety;

[0003] The existing methods for recycling and management of medical sharps mainly rely on manual identification and classification, which are inefficient and error-prone. Traditional image recognition technologies have problems of insufficient recognition accuracy and robustness when dealing with medical sharp images in complex field environments. At the same time, the existing recycling path planning methods lack consideration of the risk level and real-time status of medical sharps, and it is difficult to achieve efficient and safe recycling transportation;

[0004] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides a method and system for recycling and management of field medical sharps based on a modular anti-permeation structure, which can at least solve some problems existing in the prior art.

[0006] In the first aspect of the embodiment of the present invention, a method for recycling and management of field medical sharps based on a modular anti-permeation structure is provided, including:

[0007] Construct a medical sharp feature recognition model, collect visible light images, infrared thermal imaging images, and ultrasonic scanning images corresponding to field medical sharps and combine them into medical sharp images, perform a multi-modal image fusion algorithm based on deep hashing on the medical sharp images to obtain a fused image, extract the shape, material, texture, and internal structure corresponding to the field medical sharps from the fused image and combine them to generate medical sharp features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharp features through a combined mechanism of channel attention and spatial attention, enhance the recalibrated medical sharp features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharp features through an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of the medical sharps;

[0008] Based on the key features, the precise location of medical sharps is obtained through an improved dense prediction network to obtain location features, a hierarchical attention-guided segmentation algorithm is used to construct a refined mask for medical sharps to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of medical sharps to obtain semantic features, the location features, segmentation features and semantic features are fused and then input into a causal inference network, an improved probability graph model is used to establish the dependence relationship between features, a multi-level classifier is constructed based on the dependence relationship, and the medical sharps are classified by combining an ensemble learning strategy to obtain the corresponding category information and risk level information of the medical sharps;

[0009] Based on the category information and the risk level information, a classification and scheduling system is constructed. An improved spatio-temporal graph neural network is used to model the dynamic state representation and extract the state representations at different spatio-temporal scales. The state representations are added to a time series prediction model to determine the state evolution trend. A hierarchical reinforcement learning algorithm is combined to generate an initial scheduling strategy. The initial scheduling strategy is optimized by an improved policy gradient algorithm in a pre-set two-stage optimization network. A dynamic allocation is performed by combining an improved auction mechanism to obtain a dynamic allocation result. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeation structure. The dynamic allocation result is adjusted by an improved optimization algorithm and the adjusted dynamic allocation result is added to a path planning module. A transportation path set is generated by combining an improved ant colony algorithm and a heuristic search strategy, and an optimal recovery route is obtained by solving with a Pareto optimization algorithm.

[0010] In an alternative embodiment,

[0011] A medical sharp feature recognition model is constructed. Visible light images, infrared thermal imaging images and ultrasonic scan images corresponding to field medical sharps are collected and combined into medical sharp images. A multi-modal image fusion algorithm based on deep hashing is performed on the medical sharp images to obtain fused images. The shape, material, texture and internal structure corresponding to the field medical sharps are extracted from the fused images and combined to generate medical sharp features. A multi-scale feature pyramid is constructed through a hierarchical attention mechanism. The medical sharp features are adaptively recalibrated by combining a channel attention and a spatial attention combination mechanism. The recalibrated medical sharp features are enhanced by a deformable convolutional network and a residual connection structure. An improved adversarial training method is used to decouple the enhanced medical sharp features to obtain independent features. The importance of the independent features is evaluated and a feature selection strategy is established to obtain the key features of medical sharps, including:

[0012] Collect visible light images, infrared thermal imaging images, and ultrasonic scanning images of medical sharp objects. Respectively construct feature extraction branches for the visible light images, infrared thermal imaging images, and ultrasonic scanning images. The feature extraction branches include multiple convolutional layers, and each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function. Map the extracted features to feature vectors through a fully connected layer;

[0013] Construct a hash coding layer. Map the feature vectors to binary hash codes through a fully connected layer and a hyperbolic tangent activation function. Calculate the Hamming distance between the binary hash codes corresponding to the visible light images, infrared thermal imaging images, and ultrasonic scanning images. Screen feature-similar images based on a preset distance threshold. Use the softmax function to calculate the fusion weight coefficients of the feature-similar images. Linearly combine the weighted features to obtain fused features;

[0014] Construct a multi-layer feature pyramid network. Adjust the number of channels of the fused features through convolution. Use max pooling operations to downsample layer by layer to obtain feature maps of different scales. Calculate channel attention weights and spatial attention weights for the feature maps. The channel attention weights are calculated through global average pooling and multiple fully connected layers. The spatial attention weights are calculated through convolution operations and a sigmoid activation function. Multiply the channel attention weights and spatial attention weights to obtain a comprehensive attention map;

[0015] Construct an offset prediction branch containing multiple convolutional layers. Predict offset values for each position of the fused features. The offset values are restricted within a preset range through a hyperbolic tangent function. Adjust the sampling positions of the standard convolution according to the offset values. Use bilinear interpolation to calculate the feature values at the sampling points and perform convolution operations to obtain enhanced features. Perform residual connection between the enhanced features and the input features to obtain optimized features;

[0016] Construct a generator network and a discriminator network. The generator network adopts an encoder-decoder structure. The encoder includes multiple downsampling blocks, and each downsampling block includes convolution, instance normalization, and rectified linear unit activation. The decoder includes multiple upsampling blocks, and each upsampling block includes transposed convolution, instance normalization, and rectified linear unit activation. The discriminator network includes multiple convolutional layers and uses a leaky rectified linear unit activation function. Perform adversarial training on the optimized features through the generator network and the discriminator network to obtain independent features;

[0017] Calculate the mutual information value between the independent features and the medical sharp object category labels. Normalize the independent features. Statistically analyze the probability distributions of the independent features and the category labels. Calculate the mutual information value based on the probability distributions. Select the features with the highest mutual information value as key features.

[0018] In an alternative embodiment,

[0019] The Hamming distance between binary hash codes is calculated as shown in the following formula: Where, D ( H 1, H 2) represents the Hamming distance between the first hash code H 1 and the second hash code H 2, n represents the number of sub - blocks into which the hash code is divided, w (i) represents the i th local sensitivity weight of the sub - block, m represents the number of bits of the hash code in each sub - block, C i , j represents the correlation between the i th hash code position and the j th hash code position, represents the value of the H 1st binary bit of the first hash code i in the j th sub - block, represents the value of the H 2nd binary bit of the second hash code i in the j th sub - block. In an alternative embodiment,

[0020] Based on the key features, the medical sharp is accurately located through an improved dense prediction network to obtain location features, a hierarchical attention - guided segmentation algorithm is used to construct a refined mask for the medical sharp to obtain segmentation features, a graph convolutional network is used to analyze the component composition relationship of the medical sharp to obtain semantic features, the location features, segmentation features and semantic features are fused and then input into a causal inference network, an improved probability graph model is used to establish the dependency relationship between features, a multi - level classifier is constructed based on the dependency relationship, and an ensemble learning strategy is combined to classify the medical sharp to obtain the category information and risk level information corresponding to the medical sharp, including:

[0021] ​Construct an improved dense prediction network to locate medical sharps based on the key features. The dense prediction network includes a main branch and an auxiliary branch. The main branch contains multiple dense connection blocks, each dense connection block contains multiple convolutional units, and each convolutional unit consists of a batch normalization layer, a rectified linear unit activation function, and a convolutional layer. The auxiliary branch contains a spatial pyramid pooling module and a feature pyramid attention module. The position coordinates and confidence scores of the medical sharps are output through the dense prediction network to generate localization features.

[0022] Determine the region of interest based on the position coordinates of the localization features, and construct a hierarchical attention-guided segmentation network within the region of interest. The segmentation network includes an encoder and a decoder. The encoder uses an improved residual network structure to extract multi-level features according to the confidence scores of the localization features. The decoder performs feature fusion through skip connections based on the multi-level features, and introduces a channel attention module and a spatial attention module to enhance the features during the feature fusion process. The weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence of the localization features. At the same time, an edge-aware module is introduced to optimize the edge details of the enhanced features to generate segmentation features.

[0023] Input the segmentation features into a graph convolutional network to analyze the component composition relationship of the medical sharp. Based on the edge information of the segmentation features, construct a graph structure. In the graph structure, nodes represent different component regions, and edges represent the connection relationships between adjacent components. The graph convolutional network contains multiple graph convolutional layers, and the feature transformation parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features to generate semantic features through feature aggregation.

[0024] According to the spatial distribution of the localization features, the region segmentation results of the segmentation features, and the component relationships of the semantic features, adopt an attention-guided adaptive fusion strategy to calculate the feature weights, and input the weighted features into a causal inference network. The temporal flow of the causal inference network models the temporal dependence relationship based on the temporal evolution law of the features, and the spatial flow models the spatial dependence relationship based on the regional distribution characteristics of the features. Based on the dependence relationship output by the causal inference network, construct a probability graph model, map the dependence relationship to the node connection relationship of a directed acyclic graph, and optimize the graph structure and determine the conditional probability distribution in combination with prior knowledge.

[0025] Extract the overall contour features, surface color distribution features, and texture features of the medical sharp based on the localization features, the segmentation features, and the semantic features respectively, and combine them to obtain the global features corresponding to the medical sharp. Extract the edge detail features, hand-held part features, and functional features of the medical sharp to obtain the local features corresponding to the medical sharp.

[0026] Construct a multi - level classifier based on the output of the probability graph model, including a coarse classifier for category division and a fine classifier for fine - grained classification. The coarse classifier uses a random forest structure to divide the global features into categories, and the fine classifier uses a gradient - boosting tree structure to perform fine - grained classification on the local features. The classification results are fused by voting. The multi - level classifier is used as a base classifier, and an ensemble learning strategy is adopted to construct a classifier set. The model parameters are optimized through cross - validation, and the category information and risk level information of the medical sharp object are output.

[0027] In an alternative implementation,

[0028] The coarse classifier uses a random forest structure to divide the global features into categories, and the fine classifier uses a gradient - boosting tree structure to perform fine - grained classification on the local features. Fusing the classification results by voting includes:

[0029] Obtain the global features corresponding to the medical sharp object, perform feature normalization processing, construct an improved random forest structure containing multi - level decision units as the coarse classifier. Set a feature selection module and a classification module in the improved random forest structure. The feature selection module performs adaptive feature sampling based on feature importance scores, and the classification module constructs decision trees based on information gain and applies regularization constraints to output the basic category of the medical sharp object.

[0030] Obtain the local features corresponding to the medical sharp object, establish a cascaded gradient - boosting tree structure as the fine classifier. Divide the cascaded gradient - boosting tree structure into a feature extraction layer and a classification layer. The feature extraction layer performs attention weighting on the local features, and the classification layer performs residual learning and iteratively optimizes the tree structure parameters, implementing sample balancing and feature interaction operations to output the fine category of the medical sharp object.

[0031] Establish an adaptive multi - modal fusion mechanism, configure a dynamic weight calculation module and a decision fusion module. The dynamic weight calculation module adjusts the weight coefficient according to the classification effect, and the decision fusion module performs multi - level voting and initiates expert knowledge base arbitration when the confidence levels are inconsistent, and outputs the final classification result by combining the usage scenario and safety risk assessment.

[0032] In an alternative implementation,

[0033] Construct a classification scheduling system based on the category information and the risk level information. Model the dynamic state representation through an improved spatio-temporal graph neural network and extract the state representations at different spatio-temporal scales. Add the state representations to a time series prediction model and determine the state evolution trend. Combine a hierarchical reinforcement learning algorithm to generate an initial scheduling strategy. Optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network. Combine an improved auction mechanism for dynamic allocation to obtain a dynamic allocation result and construct a safety constraint model based on the geometric parameters and material parameters of the anti-penetration structure. Adjust the dynamic allocation result through an improved optimization algorithm and add the adjusted dynamic allocation result to a path planning module. Combine an improved ant colony algorithm and a heuristic search strategy to generate a set of transportation routes. Solve through a Pareto optimization algorithm to obtain the optimal recycling route, including:

[0034] Construct a spatio-temporal relationship graph of medical sharps. Take the medical sharps as nodes in the spatio-temporal relationship graph. The nodes contain the location information, category information, and risk level information of the medical sharps. Analyze the spatial distribution characteristics of the medical sharps in different regions by setting up multiple spatial convolutional layers. Analyze the time distribution law of the medical sharps by setting up a time convolutional layer. Input the spatial distribution characteristics and the time distribution law into a feature fusion module to generate a state representation of the medical sharps;

[0035] Input the state representation into a time series prediction model. The time series prediction model includes a time series encoder and a time series decoder. The time series encoder extracts periodic features and sudden features from the historical distribution data and historical recycling data of medical sharps according to the time order. The time series decoder predicts the distribution trend of medical sharps based on the periodic features and the sudden features. Generate a state evolution trend by weighting the historical distribution data and the historical recycling data through an attention mechanism;

[0036] Construct a hierarchical reinforcement learning network based on the state evolution trend. The hierarchical reinforcement learning network includes a high-level policy network and a low-level execution network. The high-level policy network divides the recycling tasks of medical sharps into priorities according to regions and types based on the state evolution trend to obtain a set of sub-tasks. The low-level execution network determines the recycling time and recycling route based on the set of sub-tasks to obtain an initial scheduling strategy. Store the initial scheduling strategy in an experience replay pool and select scheduling experiences through priority sampling;

[0037] Input the initial scheduling strategy into a two-stage optimization network. In the first stage of the two-stage optimization network, evaluate the value of scheduling actions through a policy gradient algorithm and introduce an entropy regularization term to adjust the policy parameters. In the second stage of the two-stage optimization network, perform resource allocation for medical sharps based on an auction mechanism and generate a dynamic allocation result according to the bidding results of recycling points and constraint conditions;

[0038] Construct a safety constraint model based on the dynamic allocation result. The safety constraint model selects a penetration-proof container according to the physical characteristics and danger levels of medical sharps, monitors the loading situation of the penetration-proof container in real time, automatically adjusts the transportation plan when detecting potential safety hazards, inputs the transportation plan into a path planning module. The path planning module constructs a path graph between collection points using the ant colony algorithm, conducts path search based on transportation distance, time window, and safety factors, and optimizes the path sequence through pheromone update and local search to generate an optimal path set that meets multi-objective constraints.

[0039] In an alternative implementation,

[0040] The second stage of the two-stage optimization network conducts resource allocation for medical sharps based on an auction mechanism. The dynamic allocation result generated according to the bidding results and constraint conditions of collection points includes:

[0041] Construct bidding entities, set medical sharp collection points as buyer entities, and set the medical sharps to be recycled as seller entities. The medical sharp collection points have a preset bidding budget and processing capacity constraints. The processing capacity constraints include the number of medical sharps that can be processed per unit time and the processing capacity limits for different types of medical sharps. Set a reserve price based on the category information, danger level, and urgency of the medical sharps to be recycled;

[0042] Adopt a multi-round increasing bidding method to construct bidding rules. The medical sharp collection points submit bids that meet the maximum bid limit, budget limit, and processing capacity requirements in each round of bidding. When a new bid is submitted by a medical sharp collection point, start a price update mechanism, and the remaining medical sharp collection points follow up with bids within a specified time, and promote the bidding process based on the minimum price increment limit;

[0043] Construct a bidding strategy optimization mechanism based on historical bidding data and the current market situation. Predict the optimal bid range by analyzing historical transaction prices, competitors' bidding patterns, and market supply and demand relationships, and set dynamic allocation constraint conditions including time window constraints, path constraints, and capacity constraints. The time window constraint is used to ensure the time limit requirements for the recycling of medical sharps, the path constraint is used to ensure the feasibility of the transportation route, and the capacity constraint is used to limit the maximum processing volume of the medical sharp collection points;

[0044] Construct a multi-objective evaluation system including an allocation efficiency dimension, a cost control dimension, and a risk management dimension. The allocation efficiency dimension includes timeliness indicators and balance indicators for the recycling of medical sharps. The cost control dimension includes transportation cost indicators and processing cost indicators. The risk management dimension includes potential safety hazard indicators and emergency response capacity indicators;

[0045] Generate a preliminary allocation plan based on the bidding result and the dynamic allocation constraint conditions, verify the feasibility of the preliminary allocation plan, and when there are constraint conflicts, generate a dynamic allocation result that meets the actual needs by adjusting the allocation quantity or replacing the medical sharp waste collection point.

[0046] In the second aspect of the embodiments of the present invention, a field medical sharp waste recycling management system based on a module anti-permeation structure is provided, including:

[0047] A first unit for constructing a medical sharp waste feature recognition model, collecting visible light images, infrared thermal imaging images, and ultrasonic scanning images corresponding to field medical sharp waste and combining them into medical sharp waste images, performing a multi-modal image fusion algorithm based on deep hashing on the medical sharp waste images to obtain a fused image, extracting the shape, material, texture, and internal structure corresponding to the field medical sharp waste from the fused image and combining them to generate medical sharp waste features, constructing a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrating the medical sharp waste features through a combined mechanism of channel attention and spatial attention, enhancing the recalibrated medical sharp waste features through a deformable convolutional network and a residual connection structure, decoupling the enhanced medical sharp waste features through an improved adversarial training method to obtain independent features, evaluating the importance of the independent features and establishing a feature selection strategy to obtain the key features of the medical sharp waste;

[0048] A second unit for precisely positioning the medical sharp waste based on the key features of the medical sharp waste through an improved dense prediction network to obtain positioning features, constructing a refined mask for the medical sharp waste through a segmentation algorithm guided by hierarchical attention to obtain segmentation features, analyzing the component composition relationship of the medical sharp waste through a graph convolutional network to obtain semantic features, fusing the positioning features, segmentation features, and semantic features and inputting them into a causal inference network, establishing a dependency relationship between features through an improved probability graph model, constructing a multi-level classifier based on the dependency relationship, and classifying the medical sharp waste in combination with an ensemble learning strategy to obtain the category information and risk level information corresponding to the medical sharp waste;

[0049] The third unit is used to construct a classification scheduling system based on the category information and the danger level information. It models the dynamic state representation through an improved spatio-temporal graph neural network and extracts the state representations at different spatio-temporal scales. The state representations are added to a time series prediction model to determine the state evolution trend. Combining with a hierarchical reinforcement learning algorithm to generate an initial scheduling strategy, the initial scheduling strategy is optimized through an improved policy gradient algorithm in a pre-set two-stage optimization network, and dynamic allocation is carried out by combining with an improved auction mechanism to obtain a dynamic allocation result. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeation structure. The dynamic allocation result is adjusted through an improved optimization algorithm and the adjusted dynamic allocation result is added to the path planning module. Combining with an improved ant colony algorithm and a heuristic search strategy to generate a set of transportation paths, and the optimal recovery route is obtained by solving through a Pareto optimization algorithm.

[0050] In the third aspect of the embodiments of the present invention,

[0051] a kind of electronic device is provided, including:

[0052] a processor;

[0053] a memory for storing instructions executable by the processor;

[0054] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0055] In the fourth aspect of the embodiments of the present invention,

[0056] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0057] In the present invention, through technical means such as multi-modal image fusion, hierarchical attention mechanism, deformable convolutional network and adversarial training, the key features of medical sharp objects are extracted, effectively improving the recognition accuracy of medical sharp objects in complex field environments, and being able to accurately identify the categories and danger levels of medical sharp objects. A classification scheduling system is constructed, and by using an improved spatio-temporal graph neural network, a time series prediction model and a hierarchical reinforcement learning algorithm, combined with an improved auction mechanism and an optimization algorithm, intelligent scheduling and dynamic allocation of medical sharp object recycling are realized, which can effectively improve the recycling efficiency and reduce the recycling cost. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeation structure, and path planning is carried out by combining with an improved ant colony algorithm and a heuristic search strategy to ensure the safety of the medical sharp object recycling process and prevent secondary pollution and accidental injuries. Description of the Drawings

[0058] Figure 1Schematic flowchart of the field medical sharp instrument recycling management method based on the module anti-permeation structure in the embodiments of the present invention;

[0059] Figure 2 Schematic structural diagram of the field medical sharp instrument recycling management system based on the module anti-permeation structure in the embodiments of the present invention. Detailed implementation manners

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0062] Figure 1 Schematic flowchart of the field medical sharp instrument recycling management method based on the module anti-permeation structure in the embodiments of the present invention, as Figure 1 shown, the method includes:

[0063] S1. Construct a medical sharp instrument feature recognition model, collect visible light images, infrared thermal imaging images, and ultrasonic scan images corresponding to the field medical sharp instruments and combine them into medical sharp instrument images, perform a multi-modal image fusion algorithm based on deep hashing on the medical sharp instrument images to obtain a fused image, extract the shape, material, texture, and internal structure corresponding to the field medical sharp instruments from the fused image and combine them to generate medical sharp instrument features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, perform adaptive recalibration on the medical sharp instrument features through a combined mechanism of channel attention and spatial attention, enhance the recalibrated medical sharp instrument features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharp instrument features through an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of the medical sharp instruments;

[0064] The multi-modal image fusion algorithm based on deep hashing is an image fusion method that combines deep learning and hashing technology. By mapping images of different modalities into a shared hashing coding space, it realizes the efficient fusion and retrieval of image information. The adaptive recalibration refers to dynamically adjusting the feature space mapping of images in multi-modal image fusion to adapt to the characteristics of different image data. The deformable convolutional network is a method to extend the traditional convolutional neural network. By introducing deformable convolution operations, it can dynamically adjust the shape and sampling position of the convolutional kernel, thus better adapting to the shape and changes of the target object, and is particularly suitable for processing image tasks with large geometric deformations, such as object detection and image segmentation. The improved adversarial training method optimizes the training process based on the generative adversarial network. Usually, by adjusting the training strategies of the generator and discriminator, it enhances the stability and effect of training.

[0065] In an alternative embodiment,

[0066] Construct a medical sharp feature recognition model. Collect visible light images, infrared thermal imaging images, and ultrasonic scan images corresponding to field medical sharps and combine them into medical sharp images. Execute the multi-modal image fusion algorithm based on deep hashing on the medical sharp images to obtain fused images. Extract the shape, material, texture, and internal structure corresponding to the field medical sharps from the fused images and combine them to generate medical sharp features. Construct a multi-scale feature pyramid through a hierarchical attention mechanism, and adaptively recalibrate the medical sharp features by combining the combined mechanism of channel attention and spatial attention. Enhance the recalibrated medical sharp features through a deformable convolutional network and a residual connection structure. Use the improved adversarial training method to decouple the enhanced medical sharp features to obtain independent features, evaluate the importance of the independent features, and establish a feature selection strategy to obtain the key features of medical sharps, including:

[0067] Collect visible light images, infrared thermal imaging images, and ultrasonic scan images of medical sharps. Respectively construct feature extraction branches for the visible light images, infrared thermal imaging images, and ultrasonic scan images. Each feature extraction branch contains multiple convolutional layers, and each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function. Map the extracted features into feature vectors through a fully connected layer;

[0068] Construct a hashing coding layer. Map the feature vectors into binary hash codes through a fully connected layer and a hyperbolic tangent activation function. Calculate the Hamming distance between the binary hash codes corresponding to the visible light images, infrared thermal imaging images, and ultrasonic scan images. Screen feature-similar images based on a preset distance threshold. Use the softmax function to calculate the fusion weight coefficients of the feature-similar images, and linearly combine the weighted features to obtain fused features;

[0069] Construct a multi - layer feature pyramid network, adjust the number of channels of the fused feature through convolution, and perform max - pooling operations layer by layer for downsampling to obtain feature maps of different scales. Calculate channel attention weights and spatial attention weights for the feature maps. The channel attention weights are calculated through global average pooling and multi - layer fully - connected layers, and the spatial attention weights are calculated through convolution operations and sigmoid activation functions. Multiply the channel attention weights and spatial attention weights to obtain a comprehensive attention map;

[0070] Construct an offset prediction branch containing multiple convolutional layers to predict offset values for each position of the fused feature. The offset values are restricted within a preset range through the hyperbolic tangent function. Adjust the sampling positions of the standard convolution according to the offset values, calculate the feature values at the sampling points using bilinear interpolation and perform convolution operations to obtain enhanced features. Perform residual connection between the enhanced features and the input features to obtain optimized features;

[0071] Construct a generator network and a discriminator network. The generator network adopts an encoder - decoder structure. The encoder contains multiple down - sampling blocks, each of which contains convolution, instance normalization, and rectified linear unit activation. The decoder contains multiple up - sampling blocks, each of which contains transposed convolution, instance normalization, and rectified linear unit activation. The discriminator network contains multiple convolutional layers and uses a leaky rectified linear unit activation function. Perform adversarial training on the optimized features through the generator network and the discriminator network to obtain independent features;

[0072] Calculate the mutual information value between the independent features and the medical sharp object category labels, perform normalization processing on the independent features, statistically analyze the probability distributions of the independent features and the category labels, calculate the mutual information value based on the probability distributions, and select the features with the highest mutual information value as key features.

[0073] The binary hash code is a special coding method in the hash algorithm, which is used for fast similarity search and retrieval by mapping data to binary codes. The Hamming distance is an index used to measure the similarity of two binary strings, representing the number of different bits in the corresponding positions of two binary strings. The hyperbolic tangent function is a commonly used activation function, and its output value ranges from -1 to 1. By compressing the output range, it helps the neural network perform non-linear transformation during training and effectively solves the problem of gradient disappearance. The leaky rectified linear unit activation function is an improved ReLU activation function that allows negative values to pass through with a certain small slope, thus avoiding the "dead neuron" problem that may occur in ReLU during training and enhancing the learning ability and stability of the model. The mutual information value is an index used to measure the degree of mutual dependence between two variables, reflecting the contribution of the information volume of one variable to another variable.

[0074] Collect visible light images, infrared thermal imaging images, and ultrasonic scanning images of the target medical sharp object. For example, use a digital camera to obtain visible light images, an infrared thermal imager to obtain infrared thermal imaging images, and a portable ultrasonic scanner to obtain ultrasonic scanning images. Each type of image reflects different characteristics of the medical sharp object and provides rich original information for subsequent feature extraction and fusion.

[0075] Construct feature extraction branches for the three collected images respectively. Each branch contains multiple convolutional layers, and after each convolutional layer, a batch normalization layer and a rectified linear unit activation function are connected. The convolutional layer is used to extract local features of the image. The batch normalization layer can accelerate model training and improve the generalization ability of the model. The rectified linear unit activation function introduces non-linearity and enhances the expression ability of the model. Taking the visible light image as an example, assume the input image size is 256x256 pixels. The first convolutional layer uses 32 convolutional kernels with a size of 3x3, a stride of 1, and a padding of 1. Then the output feature map size is still 256x256, and the number of channels becomes 32. The subsequent convolutional layers can adjust parameters such as the number, size, stride, and padding of the convolutional kernels according to needs. Finally, the extracted features are mapped into feature vectors through a fully connected layer. For example, the 256x256x32 feature map is converted into a 1024-dimensional feature vector.

[0076] Hash encode and fuse the feature vectors of three types of images. Map the feature vectors to binary hash codes, such as 128-bit hash codes, through a fully connected layer and a hyperbolic tangent activation function. Calculate the Hamming distance between the binary hash codes corresponding to the three types of images. The smaller the Hamming distance, the more similar the image features are. Set a distance threshold, such as 32, to filter out images with similar features. Then, use the softmax function to calculate the fusion weight coefficients of the images with similar features. For example, assume that the Hamming distance between the visible light image and the infrared thermal imaging image is 20, and the Hamming distance between the visible light image and the ultrasonic scan image is 40. Then, the visible light image and the infrared thermal imaging image are more similar, and their corresponding weight coefficients are also larger. Finally, linearly combine the weighted features to obtain the fused features. The fused features integrate the information of the three types of images and more comprehensively describe the characteristics of the medical sharp object.

[0077] After obtaining the fused features, construct a multi-layer feature pyramid network. Adjust the number of channels of the fused features through convolution, and then perform max-pooling operations for downsampling layer by layer to obtain feature maps of different scales. For example, the size of the initial fused feature is 64x64, and through max-pooling operations, feature maps of different scales such as 32x32 and 16x16 can be obtained. Feature maps of different scales can capture feature information of different sizes. Calculate the channel attention weights and spatial attention weights for each scale of the feature maps. The channel attention weights are calculated through global average pooling and multi-layer fully connected operations, and the spatial attention weights are calculated through convolution operations and a sigmoid activation function. Multiply the channel attention weights and the spatial attention weights to obtain a comprehensive attention map for adaptively recalibrating the features and highlighting important feature information.

[0078] Enhance the recalibrated features. Construct an offset prediction branch containing multiple convolutional layers to predict the offset values for each position of the fused features. The offset values are restricted within a preset range, such as [-1, 1], through the hyperbolic tangent function. Adjust the sampling positions of the standard convolution according to the offset values, calculate the feature values at the sampling points through bilinear interpolation and perform convolution operations to obtain the enhanced features. Connect the enhanced features with the input features through a residual connection to obtain the optimized features, further improving the expression ability of the features.

[0079] Decouple and select features from the optimized features. Construct a generator network and a discriminator network, and decouple the optimized features through adversarial training to obtain independent features. Calculate the mutual information value between the independent features and the medical sharp object category labels. Normalize the independent features, statistically analyze the probability distributions of the independent features and the category labels, and calculate the mutual information value based on the probability distributions. Select the features with higher mutual information values as the key features for the final medical sharp object recognition.

[0080] In this embodiment, through multimodal image fusion and deep learning techniques, more comprehensive and discriminative features can be extracted, thereby improving the accuracy of medical sharp object recognition. It has strong robustness to environmental factors such as illumination, angle, and occlusion, and can adapt to complex field environments. Through the feature selection strategy, redundant features can be removed, the computational complexity can be reduced, and the recognition efficiency can be improved.

[0081] In an alternative embodiment,

[0082] The Hamming distance between binary hash codes is calculated as shown in the following formula: Where, D ( H 1, H 2) represents the first hash code H 1 and the second hash code H 2, n represents the number of sub-blocks into which the hash code is divided, w (i) represents the i -th local sensitivity weight of the sub-block, m represents the number of bits of the hash code in each sub-block, C i , j represents the correlation between the i -th hash code position and the j -th hash code position, represents the value of the H -th binary bit of the first hash code i 1 in the j -th sub-block, represents the value of the H -th binary bit of the second hash code i 2 in the j -th sub-block.

[0083] In this embodiment, through the weighted Hamming distance, similar and dissimilar data can be more effectively distinguished, thereby improving the retrieval efficiency and finding the target data faster. The locality-sensitive hashing and the weighting mechanism can reduce the influence of noise and errors, improve the robustness of the retrieval, and obtain better retrieval results even when there is a certain amount of noise or error in the data. By assigning different weights to different sub-blocks, the similarity between data can be more accurately reflected, thereby improving the retrieval accuracy and finding more relevant results.

[0084] ​S2. Based on the key features, precisely locate the medical sharp objects through an improved dense prediction network to obtain location features, construct a refined mask for the medical sharp objects using a hierarchical attention-guided segmentation algorithm to obtain segmentation features, analyze the component composition relationship of the medical sharp objects using a graph convolutional network to obtain semantic features, fuse the location features, segmentation features, and semantic features and input them into a causal inference network, establish the dependency relationship between features through an improved probabilistic graph model, construct a multi-level classifier based on the dependency relationship, and classify the medical sharp objects in combination with an ensemble learning strategy to obtain the category information and risk level information corresponding to the medical sharp objects;

[0085] The improved dense prediction network is a deep learning model for image segmentation and object detection tasks. By enhancing the feature extraction ability and fine-grained prediction of the network, the performance of the traditional dense prediction network is improved. The hierarchical attention-guided segmentation algorithm is an image segmentation method based on the attention mechanism. By applying hierarchical attention guidance to feature maps at different levels, the model can better focus on important regions in the image, thereby improving the segmentation accuracy. The causal inference network is a deep learning model based on causal inference, used to infer the interaction between variables in complex causal relationships. The improved probabilistic graph model is an enhanced graph model that combines the ideas of probability theory and graph theory, usually used to describe the dependency relationship between random variables.

[0086] In an alternative implementation,

[0087] Based on the key features, precisely locate the medical sharp objects through an improved dense prediction network to obtain location features, construct a refined mask for the medical sharp objects using a hierarchical attention-guided segmentation algorithm to obtain segmentation features, analyze the component composition relationship of the medical sharp objects using a graph convolutional network to obtain semantic features, fuse the location features, segmentation features, and semantic features and input them into a causal inference network, establish the dependency relationship between features through an improved probabilistic graph model, construct a multi-level classifier based on the dependency relationship, and classify the medical sharp objects in combination with an ensemble learning strategy to obtain the category information and risk level information corresponding to the medical sharp objects, including:

[0088] Construct an improved dense prediction network, locate the medical sharp objects based on the key features. The dense prediction network includes a main branch and an auxiliary branch. The main branch contains multiple dense connection blocks, each dense connection block contains multiple convolutional units, and each convolutional unit consists of a batch normalization layer, a rectified linear unit activation function, and a convolutional layer. The auxiliary branch contains a spatial pyramid pooling module and a feature pyramid attention module. Output the position coordinates and confidence scores of the medical sharp objects through the dense prediction network to generate location features;

[0089] Determine the region of interest based on the position coordinates of the positioning features, and construct a hierarchical attention-guided segmentation network within the region of interest. The segmentation network includes an encoder and a decoder. The encoder extracts multi-level features using an improved residual network structure according to the confidence score of the positioning features. The decoder performs feature fusion through skip connections based on the multi-level features, and introduces a channel attention module and a spatial attention module to enhance the features during the feature fusion process. The weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence of the positioning features. At the same time, an edge perception module is introduced to optimize the edge details of the enhanced features to generate segmentation features;

[0090] Input the segmentation features into a graph convolutional network to analyze the component composition relationship of the medical sharp instrument. Based on the edge information of the segmentation features, construct a graph structure, where the nodes in the graph structure represent different component regions, and the edges represent the connection relationships between adjacent components. The graph convolutional network contains multiple graph convolutional layers, and the feature transformation parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation;

[0091] According to the spatial distribution of the positioning features, the regional segmentation results of the segmentation features, and the component relationship of the semantic features, adopt an attention-guided adaptive fusion strategy to calculate the feature weights, and input the weighted features into a causal inference network. The temporal flow of the causal inference network models the temporal dependence relationship based on the temporal evolution law of the features, and the spatial flow models the spatial dependence relationship based on the regional distribution characteristics of the features. Based on the dependence relationship output by the causal inference network, construct a probability graph model, map the dependence relationship to the node connection relationship of a directed acyclic graph, and optimize the graph structure and determine the conditional probability distribution in combination with prior knowledge;

[0092] Extract the overall contour feature, surface color distribution feature, and texture feature of the medical sharp instrument based on the positioning feature, the segmentation feature, and the semantic feature respectively, and combine them to obtain the global feature corresponding to the medical sharp instrument. Extract the edge detail feature, hand-held part feature, and functional feature of the medical sharp instrument to obtain the local feature corresponding to the medical sharp instrument;

[0093] Construct a multi-level classifier according to the output of the probability graph model, including a coarse classifier for category division and a fine classifier for fine classification. Among them, the coarse classifier uses a random forest structure to classify the global features, and the fine classifier uses a gradient boosting tree structure to perform fine classification on the local features. The classification results are fused by voting. Take the multi-level classifier as the base classifier, adopt an ensemble learning strategy to construct a classifier set, optimize the model parameters through cross-validation, and output the category information and danger level information of the medical sharp instrument.

[0094] The edge perception module is a component in a deep neural network, specifically designed to improve the sensitivity of the model when processing image edge information. The attention-guided adaptive fusion strategy is a strategy that weights and fuses features of different modalities or different levels through an attention mechanism. The temporal flow refers to the propagation or change process of data in a time series. The gradient boosting tree structure is an ensemble learning method that gradually improves the prediction performance of the model through the combination of multiple decision trees.

[0095] Construct an improved dense prediction network to accurately locate medical sharps, which includes a main branch and an auxiliary branch. The main branch consists of multiple densely connected blocks, and each block contains several convolutional units. Each convolutional unit sequentially includes a batch normalization layer, a ReLU activation function layer, and a convolutional layer. The auxiliary branch includes a spatial pyramid pooling module and a feature pyramid attention module. After inputting an image of a medical sharp, the main branch extracts multi-scale features of the image, while the auxiliary branch captures the global context information and local significant features of the image. The output features of the main branch and the auxiliary branch are fused to obtain the position coordinates and confidence scores of the sharp, generating localization features. For example, when inputting an image containing a scalpel, the network outputs the coordinates of the scalpel tip (x = 100, y = 200) and a confidence score of 0.95.

[0096] Determine the region of interest based on the position coordinates of the localization features, and construct a hierarchical attention-guided segmentation network within the region of interest, which includes an encoder and a decoder. The encoder, based on the confidence scores of the localization features, uses an improved residual network structure to extract multi-level features. The decoder then fuses the multi-level features extracted by the encoder through skip connections. During the feature fusion process, a channel attention module and a spatial attention module are introduced to enhance the features, and the weights of these two modules are adaptively adjusted according to the confidence scores of the localization features. At the same time, an edge perception module is introduced to optimize the edge details of the enhanced features, finally generating segmentation features. For example, based on the coordinates and confidence of the scalpel tip, determine the region of interest containing the entire scalpel, and the segmentation network outputs a pixel-level mask of the scalpel.

[0097] Input the segmentation features into a graph convolutional network to analyze the component composition relationship of medical sharps. Based on the edge information of the segmentation features, construct a graph structure, where the nodes in the graph represent different component regions, and the edges represent the connection relationships between adjacent components. The graph convolutional network includes multiple graph convolutional layers, and the feature transformation parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, generating semantic features through feature aggregation. For example, the segmentation result of a scalpel can construct a graph, with the handle and the blade as nodes and their connection relationship as the edge, and the graph convolutional network analyzes to obtain the semantic features of the handle and the blade.

[0098] According to the spatial distribution of the positioning features, the regional segmentation results of the segmentation features, and the part relationships of the semantic features, an attention-guided adaptive fusion strategy is adopted to calculate the feature weights, and the weighted features are input into the causal inference network. The temporal flow of the causal inference network models the temporal dependence relationship based on the temporal evolution law of the features, and the spatial flow models the spatial dependence relationship based on the regional distribution characteristics of the features. A probabilistic graphical model is constructed based on the dependence relationship output by the causal inference network, mapping the dependence relationship to the node connection relationship of a directed acyclic graph, and optimizing the graph structure and determining the conditional probability distribution in combination with prior knowledge. For example, the dependence relationship between the tip position, blade shape, and handle material of a scalpel is modeled into a probabilistic graphical model.

[0099] Based on the positioning features, segmentation features, and semantic features, the overall contour features, surface color distribution features, and texture features of the medical sharp instrument are extracted respectively, and the global features are obtained by combination. Then, the edge detail features, hand-held part features, and functional features of the medical sharp instrument are extracted to obtain local features. For example, the overall shape, color, and texture of a scalpel are extracted as global features, and the sharpness of the blade, the holding method of the handle, and the cutting function of the scalpel are extracted as local features.

[0100] A multi-level classifier is constructed according to the output of the probabilistic graphical model, including a coarse classifier for class division and a fine classifier for fine classification. The coarse classifier uses a random forest structure to classify the global features, the fine classifier uses a gradient boosting tree structure to perform fine classification on the local features, and the classification results are fused by voting. The multi-level classifier is used as the base classifier, and an ensemble learning strategy is adopted to construct a classifier set, and the model parameters are optimized through cross-validation, and the category information and danger level information of the medical sharp instrument are output. For example, the coarse classifier classifies the scalpel as a cutting tool category, the fine classifier further determines it as a scalpel, and evaluates its danger level as high according to the sharpness of the blade.

[0101] In this embodiment, through multi-feature fusion and causal inference, the category and danger level of the medical sharp instrument can be identified more accurately, the mis-identification rate can be reduced, potential medical accidents can be effectively avoided, and technologies such as a dense prediction network, a hierarchical attention mechanism, and a graph convolutional network are combined, which can effectively handle complex situations such as image noise, illumination changes, and occlusion, improve the robustness of the recognition, can finely segment the medical sharp instrument, analyze its component composition relationship, and extract global and local features, so as to achieve more refined recognition and danger level assessment, and provide a more reliable guarantee for medical safety.

[0102] In an alternative embodiment,

[0103] The coarse classifier uses a random forest structure to classify the global features, the fine classifier uses a gradient boosting tree structure to perform fine classification on the local features, and the classification results are fused by voting, including:

[0104] Obtain the global features corresponding to the medical sharp instrument, perform feature normalization processing, construct an improved random forest structure containing multi-level decision units as a coarse classifier, set a feature selection module and a classification module in the improved random forest structure, perform adaptive feature sampling by the feature selection module based on the feature importance score, use the classification module to construct a decision tree based on information gain and impose regularization constraints, and output the basic category of the medical sharp instrument;

[0105] Obtain the local features corresponding to the medical sharp instrument, establish a cascaded gradient boosting tree structure as a fine classifier, divide the cascaded gradient boosting tree structure into a feature extraction layer and a classification layer, perform attention weighting on the local features by the feature extraction layer, perform residual learning by the classification layer and iteratively optimize the tree structure parameters, implement sample balancing and feature interaction operations, and output the fine category of the medical sharp instrument;

[0106] Establish an adaptive multi-modal fusion mechanism, configure a dynamic weight calculation module and a decision fusion module, use the dynamic weight calculation module to adjust the weight coefficient according to the classification effect, perform multi-level voting by the decision fusion module and initiate expert knowledge base arbitration when the confidence levels are inconsistent, and output the final classification result by combining the usage scenario and safety risk assessment.

[0107] The expert knowledge base arbitration is a decision-making mechanism that combines artificial intelligence and expert systems and is usually used to resolve different opinions or information conflicts from multiple sources. In the system, the expert knowledge base contains the knowledge and rules of domain experts, and the arbitration process makes decisions based on the knowledge in the expert library to help the system make a final choice or judgment, and is often used in intelligent decision-making in fields such as medicine, law, and finance.

[0108] Obtain the global features of the medical sharp instrument. The global features describe the overall shape and structural information of the medical sharp instrument, such as the length, width, thickness, volume, surface area, etc. of the sharp instrument. Taking a disposable syringe needle as an example, its global features can be expressed as: length is 38 mm, width is 0.5 mm, thickness is 0.3 mm, volume is 0.057 cubic millimeters, and surface area is 38.5 square millimeters. Perform normalization processing on the obtained global features to scale each feature value to the same numerical range, such as between 0 and 1, to eliminate the influence of the dimension between different features.

[0109] Construct an improved random forest structure as the coarse classifier. This structure contains multi-level decision units, and each unit consists of a feature selection module and a classification module. The feature selection module performs adaptive feature sampling based on feature importance scores. For example, in the classification of needles, if the importance score of the "length" feature is higher than that of the "surface area" feature, the feature selection module will select the "length" feature with a higher probability to participate in the construction of the decision tree. The classification module constructs a decision tree based on information gain and imposes regularization constraints, such as restricting the maximum depth of the decision tree or the minimum number of samples in the leaf nodes, to prevent overfitting. The normalized global features are classified by the improved random forest structure, and the basic categories of medical sharps are output, such as needles, scalpels, suture needles, etc. Suppose that according to the global features, the needle is preliminarily classified as the "syringe needle" category.

[0110] Obtain the local features of the medical sharp. The local features describe the detailed information of specific parts of the medical sharp, such as the shape of the needle tip, the sharpness of the blade, the curvature of the suture needle, etc. The local features of the needle can include: the needle tip angle, the needle tip curvature, the texture of the needle body, etc. Taking the needle tip angle as an example, its value is 30 degrees.

[0111] Establish a cascaded gradient boosting tree structure as the fine classifier. This structure is divided into a feature extraction layer and a classification layer. The feature extraction layer performs attention weighting on the local features. For example, the needle tip angle has higher importance in distinguishing different types of syringe needles, so the feature extraction layer will assign a higher weight to this feature. The classification layer performs residual learning and iteratively optimizes the tree structure parameters, such as the learning rate, the number and depth of the trees, etc. At the same time, sample balancing and feature interaction operations are implemented to improve the robustness and generalization ability of the classifier. Suppose that the local features of the needle indicate that its needle tip angle is small and suitable for subcutaneous injection, then the fine classifier further classifies it as a "subcutaneous injection needle".

[0112] Establish an adaptive multi-modal fusion mechanism, which includes a dynamic weight calculation module and a decision fusion module. The dynamic weight calculation module adjusts the weight coefficients according to the classification effects of the coarse classifier and the fine classifier. For example, if the accuracy of the fine classifier is high, a higher weight is assigned to it. The decision fusion module performs multi-level voting, such as voting separately according to the coarse classification result and the fine classification result, and determines the final classification result according to the voting result. When the confidence levels are inconsistent, for example, the coarse classification result is "syringe needle" and the fine classification result is "suture needle", then the expert knowledge base arbitration is initiated, and the final classification result is output by combining the usage scenario and the safety risk assessment. Suppose that the medical sharp is finally determined to be a "subcutaneous injection needle".

[0113] In this embodiment, by combining global and local features, as well as the strategies of coarse classification and fine classification, the characteristic information of medical sharps can be captured more comprehensively, thereby improving the accuracy of classification. The improved random forest and cascade gradient boosting tree structures can quickly process a large amount of feature data, thereby improving the classification efficiency. Through accurate classification, the classification and treatment of medical waste can be effectively guided, the risk of medical waste treatment can be reduced, and the safety of the medical environment can be guaranteed.

[0114] S3. Based on the category information and the risk level information, construct a classification and scheduling system. Model the dynamic state representation through an improved spatio-temporal graph neural network and extract the state representations at different spatio-temporal scales. Add the state representations to the time series prediction model and determine the state evolution trend. Combine the hierarchical reinforcement learning algorithm to generate an initial scheduling strategy. Optimize the initial scheduling strategy through the improved policy gradient algorithm in the pre-set two-stage optimization network. Combine the improved auction mechanism for dynamic allocation to obtain the dynamic allocation result. Based on the geometric parameters and material parameters of the anti-permeation structure, construct a safety constraint model. Adjust the dynamic allocation result through the improved optimization algorithm and add the adjusted dynamic allocation result to the path planning module. Combine the improved ant colony algorithm and heuristic search strategy to generate a set of transportation paths, and solve to obtain the optimal recovery route through the Pareto optimization algorithm.

[0115] The improved spatio-temporal graph neural network is an enhanced model that combines graph neural networks with time series modeling, used to capture the spatio-temporal dependence relationships between nodes. The state evolution trend refers to the dynamic change law of a system or variable in the time dimension, usually used to analyze time series data or predict the change trend of future states. The two-stage optimization network is a deep learning model trained in stages, usually including a coarse-grained optimization stage and a fine-grained optimization stage. The improved auction mechanism is an optimized resource allocation method that improves the fairness and revenue maximization effect of the auction process by introducing more efficient bidding rules or more reasonable allocation strategies, and is widely used in fields such as e-commerce and resource scheduling. The safety constraint model is a mathematical model that introduces safety constraints in an optimization problem, and helps decision-makers optimize the objective function while ensuring that the system operates without violating safety conditions. The ant colony algorithm is an optimization algorithm that simulates the foraging behavior of ants and collaboratively completes the search for the global optimal path through the pheromone transmission between individuals, and is often used to solve complex combinatorial optimization problems. The heuristic search strategy is an algorithm strategy that uses heuristic information to guide the search direction, and significantly improves the search efficiency by selecting paths or nodes that are more likely to reach the goal, and is common in the fields of artificial intelligence and operations research optimization.

[0116] In an alternative embodiment,

[0117] Construct a classification and scheduling system based on the category information and the risk level information. Model the dynamic state representation through an improved spatio-temporal graph neural network and extract the state representation at different spatio-temporal scales. Add the state representation to the time series prediction model and determine the state evolution trend. Combine the hierarchical reinforcement learning algorithm to generate an initial scheduling strategy. Optimize the initial scheduling strategy through the improved policy gradient algorithm in the pre-set two-stage optimization network. Combine the improved auction mechanism for dynamic allocation, obtain the dynamic allocation result, and construct a safety constraint model based on the geometric parameters and material parameters of the anti-penetration structure. Adjust the dynamic allocation result through the improved optimization algorithm and add the adjusted dynamic allocation result to the path planning module. Combine the improved ant colony algorithm and heuristic search strategy to generate a set of transportation paths. Solve the optimal recovery route through the Pareto optimization algorithm, including:

[0118] Construct a spatio-temporal relationship graph of medical sharps. Take the medical sharps as the nodes in the spatio-temporal relationship graph. The nodes contain the location information, category information, and risk level information of the medical sharps. Analyze the spatial distribution characteristics of the medical sharps in different regions by setting multiple spatial convolutional layers. Analyze the time distribution law of the medical sharps by setting a time convolutional layer. Input the spatial distribution characteristics and the time distribution law into the feature fusion module to generate the state representation of the medical sharps;

[0119] Input the state representation into the time series prediction model. The time series prediction model includes a time series encoder and a time series decoder. The time series encoder extracts the periodic features and sudden features from the historical distribution data and historical recovery data of the medical sharps according to the time order. The time series decoder predicts the distribution trend of the medical sharps based on the periodic features and the sudden features. Generate the state evolution trend by weighting the historical distribution data and the historical recovery data through the attention mechanism;

[0120] Construct a hierarchical reinforcement learning network based on the state evolution trend. The hierarchical reinforcement learning network includes a high-level policy network and a low-level execution network. The high-level policy network divides the recovery tasks of the medical sharps into priorities according to regions and types based on the state evolution trend to obtain a set of sub-tasks. The low-level execution network determines the recovery time and recovery route based on the set of sub-tasks to obtain an initial scheduling strategy. Store the initial scheduling strategy in the experience replay pool and select scheduling experiences through the priority sampling method;

[0121] Input the initial scheduling strategy into the two-stage optimization network. In the first stage of the two-stage optimization network, evaluate the value of the scheduling action through the policy gradient algorithm and introduce an entropy regularization term to adjust the policy parameters. In the second stage of the two-stage optimization network, perform resource allocation for the medical sharps based on the auction mechanism, and generate a dynamic allocation result according to the bidding results of the recovery points and the constraint conditions;

[0122] Construct a safety constraint model based on the dynamic allocation result. The safety constraint model selects a penetration-proof container according to the physical characteristics and risk levels of medical sharps, monitors the loading situation of the penetration-proof container in real time, automatically adjusts the transportation plan when a safety hazard is detected, inputs the transportation plan into a path planning module, and the path planning module uses the ant colony algorithm to construct a path graph between collection points, searches for paths based on transportation distance, time window, and safety factors, optimizes the path sequence through pheromone update and local search, and generates an optimal path set that meets multi-objective constraints.

[0123] The high-level policy network is a deep learning network for making global decisions. By comprehensively analyzing global state information, it provides guiding policies and target directions for the low-level network. The low-level execution network is a deep learning network focusing on specific task execution, responsible for completing fine-grained operations and feedback according to the instructions of the high-level policy, realizing the separation of policy and execution. The pheromone is the core parameter in the ant colony algorithm that simulates the behavior of ant colonies and is used to represent the quality of paths. The pheromone concentration on a path increases as more ants choose that path, thus guiding subsequent search directions.

[0124] Construct a spatio-temporal relationship graph of medical sharps. Each medical sharp is regarded as a node in the graph. Each node contains the geographical location (such as longitude and latitude), the category it belongs to (such as needles, scalpels), and the risk level (such as high-risk, low-risk) of the medical sharp. To analyze the spatial distribution characteristics of medical sharps in different regions, use multi-layer spatial convolutional layers to process the node information. Each layer of spatial convolutional layer considers the information of the node and its neighbor nodes, and aggregates this information through learnable weight parameters to extract spatial features at different scales. At the same time, to analyze the time distribution law of medical sharps, use a temporal convolutional layer to process the node information. The temporal convolutional layer considers the information of the node at different time steps and captures the changing trend in time through learnable weight parameters. Finally, input the spatial distribution features extracted by the spatial convolutional layer and the time distribution law extracted by the temporal convolutional layer into a feature fusion module, which integrates the two features to generate a state representation of the medical sharp. For example, a certain node represents a high-risk needle, its location is in area A, and the time is October 1st. Through the spatio-temporal graph neural network, the spatio-temporal distribution characteristics of this needle among high-risk sharps in area A can be obtained.

[0125] The state representation of medical sharps is input into a time series prediction model, which consists of a time series encoder and a time series decoder. The time series encoder receives the historical distribution data of medical sharps (such as the number of sharps in each area every day in the past year) and the historical recycling data (such as the number of recycled sharps in each area every day in the past year), and extracts features from these data in chronological order. The encoder will learn the periodic features (such as more sharps on Mondays) and sudden features (such as abnormal number of sharps on a certain holiday) in the data. The time series decoder then predicts the distribution trend of medical sharps in the future period based on the periodic and sudden features extracted by the encoder. To better capture the important information in the data, an attention mechanism is used in the model to weight the historical distribution data and historical recycling data, so as to generate a more accurate state evolution trend. For example, based on the data of the past year, the model predicts that the number of high-risk needles in area A will increase by 20% in the next week.

[0126] A hierarchical reinforcement learning network is constructed based on the state evolution trend, which consists of a high-level policy network and a low-level execution network. The high-level policy network divides the priority of the medical sharp recycling task according to the state evolution trend (such as the number of high-risk needles in area A will increase by 20% in the next week), and decomposes it into multiple subtasks according to the area and type. For example, "recycle high-risk needles in area A" is taken as a high-priority subtask. The low-level execution network then determines the recycling time and recycling route for each subtask based on the subtask set, so as to generate an initial scheduling strategy. For example, the low-level network decides to recycle high-risk needles in area A between 9 am and 10 am on Tuesday and plans a specific recycling route. The generated initial scheduling strategy is stored in the experience replay pool, and scheduling experiences are selected through priority sampling for subsequent optimization.

[0127] The initial scheduling strategy is input into a two-stage optimization network. In the first stage, the policy gradient algorithm is used to evaluate the value of the scheduling action, and an entropy regularization term is introduced to adjust the policy parameters to make the policy more stable and robust. In the second stage, the resources of medical sharps are allocated based on the auction mechanism. Each recycling point submits a bid according to its own situation and the current scheduling strategy, and the network generates a dynamic allocation result according to the bid results of the recycling points and the constraint conditions (such as vehicle capacity, time window). For example, recycling point B wins the bid at the lowest price and obtains the task of recycling high-risk needles in area A.

[0128] Build a safety constraint model based on the dynamic allocation results. This model selects appropriate impermeable containers according to the physical characteristics (such as size, weight) and hazard levels of medical sharps. During transportation, the loading conditions of the impermeable containers are monitored in real time, for example, by sensors to monitor parameters such as temperature and pressure inside the containers. When safety hazards are detected (such as container breakage, too high temperature), the transportation plan is automatically adjusted. The adjusted transportation plan is input into the path planning module. This module uses the ant colony algorithm to construct a path graph between collection points and conducts path search based on transportation distance, time window, and safety factors. The path sequence is optimized through techniques such as pheromone update and local search, and finally an optimal path set that meets multi-objective constraints is generated. For example, if the temperature inside the container is detected to be too high, the system will automatically adjust the transportation plan, select a shorter and safer path, and reduce the transportation speed.

[0129] In this embodiment, through the spatio-temporal graph neural network and the time series prediction model, the distribution trend and state evolution of medical sharps can be accurately predicted, so as to plan the collection tasks in advance, optimize resource allocation, and improve collection efficiency. Through the safety constraint model and impermeable containers, the safety risks during the transportation of medical sharps can be effectively reduced, preventing accidents. Through hierarchical reinforcement learning and the auction mechanism, the collection routes and resource allocation can be optimized, reducing transportation costs and improving economic benefits.

[0130] In an alternative embodiment,

[0131] The second stage of the two-stage optimization network conducts resource allocation for medical sharps based on the auction mechanism, and the dynamic allocation results generated according to the bidding results and constraint conditions of collection points include:

[0132] Construct bidding entities, set the medical sharp collection points as buyer entities, and set the medical sharps to be collected as seller entities. The medical sharp collection points have preset bidding budgets and processing capacity constraints. The processing capacity constraints include the number of medical sharps that can be processed per unit time and the processing capacity limits for different types of medical sharps. Set the reserve price based on the category information, hazard level, and urgency of the medical sharps to be collected;

[0133] Adopt a multi-round ascending bidding method to construct bidding rules. The medical sharp collection points submit bids that meet the maximum bid limit, budget limit, and processing capacity requirements in each round of bidding. When a new bid is submitted by a medical sharp collection point, start the price update mechanism, and the remaining medical sharp collection points follow up with bids within the specified time, and promote the bidding process based on the minimum price increment limit;

[0134] Build a bidding strategy optimization mechanism based on historical bidding data and the current market situation. Predict the optimal bidding range by analyzing historical transaction prices, competitors' bidding patterns, and market supply and demand relationships. Set dynamic allocation constraint conditions including time window constraints, path constraints, and capacity constraints. The time window constraint is used to ensure the time limit requirement for the completion of the recycling of medical sharps. The path constraint is used to ensure the feasibility of the transportation route. The capacity constraint is used to limit the maximum processing volume of the medical sharp recycling point.

[0135] Build a multi-objective evaluation system including the dimensions of allocation efficiency, cost control, and risk management. The allocation efficiency dimension includes timeliness indicators and balance indicators for the recycling of medical sharps. The cost control dimension includes transportation cost indicators and processing cost indicators. The risk management dimension includes safety hazard indicators and emergency response capacity indicators.

[0136] Generate a preliminary allocation plan based on the bidding results and the dynamic allocation constraint conditions, and conduct a feasibility verification on the preliminary allocation plan. When there are constraint conflicts, generate a dynamic allocation result that meets the actual needs by adjusting the allocation quantity or replacing the medical sharp recycling point.

[0137] The bidding entity is a participant in the auction mechanism, responsible for bidding on the target item, usually representing different interest parties. The reserve price is the lowest acceptable price set in the auction mechanism to ensure that the target item is not sold at too low a price. The multi-round ascending bidding method is an auction rule. By gradually increasing the bid in multiple rounds of bidding, it ensures the fairness and transparency of the bidding process and helps to maximize the bidding revenue.

[0138] Build a bidding entity. Set the medical sharp recycling point as the buyer and the medical sharps to be recycled as the seller. Each recycling point has a preset bidding budget and processing capacity constraints. The processing capacity constraints include the number of medical sharps that can be processed per unit time and the processing capacity limits for different types of medical sharps. For example, the daily processing capacity of Recycling Point A is 1000 pieces, of which the processing capacity for needle-type medical sharps is 800 pieces, and the processing capacity for blade-type medical sharps is 200 pieces, with a total budget of 5000 yuan. The daily processing capacity of Recycling Point B is 1500 pieces, of which the processing capacity for needle-type medical sharps is 500 pieces, and the processing capacity for blade-type medical sharps is 1000 pieces, with a total budget of 7000 yuan. Set the reserve price according to the category information, risk level, and urgency of the medical sharps to be recycled. For example, the reserve price for needle-type medical sharps is 1 yuan per piece, and the reserve price for blade-type medical sharps is 2 yuan per piece. If the urgency is relatively high, the reserve price will increase by 20%.

[0139] The bidding rules are constructed by adopting a multi-round incremental bidding method. Each recycling point submits a bid that meets the highest bid limit, budget limit, and processing capacity requirements in each round of bidding. For example, Recycling Point A bids 1.2 yuan per piece for a batch of 100 needle-type medical sharps, with a total price of 120 yuan, which does not exceed its budget and processing capacity. When a recycling point submits a new bid, the price update mechanism is initiated, and the remaining recycling points follow up with their bids within a specified time, such as 5 minutes. The system sets a minimum price increment, such as 0.1 yuan per piece, to promote the bidding process.

[0140] The bidding strategy optimization mechanism is constructed based on historical bidding data and the current market situation. The optimal bidding range is predicted by analyzing historical transaction prices, competitors' bidding patterns, and market supply and demand relationships. For example, the system analyzes that the average transaction price of needle-type medical sharps in the past month was 1.1 yuan per piece, and the current market demand is high, and it is recommended that Recycling Point A's bidding range be 1.15 - 1.25 yuan per piece. At the same time, dynamic allocation constraint conditions are set, including time window constraints, path constraints, and capacity constraints. For example, it is required that medical sharps be recycled within 24 hours, the transportation route must avoid sensitive areas such as schools, and the maximum processing volume of each recycling point cannot exceed its daily processing capacity.

[0141] A multi-objective evaluation system is constructed, including the allocation efficiency dimension, cost control dimension, and risk management dimension. The allocation efficiency dimension includes indicators of the timeliness and balance of medical sharp recycling. For example, all medical sharps are recycled within the specified time, and the processing volumes of each recycling point are relatively balanced. The cost control dimension includes transportation cost indicators and processing cost indicators. For example, the shortest transportation route is selected to reduce transportation costs, and the processing cost of each recycling point is controlled within its budget. The risk management dimension includes safety hazard indicators and emergency response capacity indicators. For example, necessary safety measures are taken during transportation to ensure the safe transportation of medical sharps, and an emergency plan is formulated to deal with emergencies.

[0142] A preliminary allocation plan is generated based on the bidding results and dynamic allocation constraint conditions. The feasibility of the preliminary allocation plan is verified. When there are constraint conflicts, a dynamic allocation result that meets the actual needs is generated by adjusting the allocation quantity or replacing the recycling point. For example, in the preliminary plan, 500 needle-type medical sharps are allocated to Recycling Point A, but the remaining processing capacity of Recycling Point A for needle-type medical sharps on that day is only 400 pieces. Then, the remaining 100 pieces need to be allocated to other recycling points or the allocation quantity of Recycling Point A needs to be adjusted.

[0143] In this embodiment, it is possible to quickly match the medical sharp instrument recycling demand with the processing capacity of the recycling point, realize the timeliness and balance of medical sharp instrument recycling, improve the recycling efficiency. Through the bidding mechanism and dynamic allocation constraints, it is possible to effectively control the transportation cost and processing cost, achieve cost optimization, consider risk factors such as potential safety hazards and emergency response capabilities, effectively manage risks, and ensure the safety and reliability of the medical sharp instrument recycling process.

[0144] Figure 2 FIG. is a schematic structural diagram of a field medical sharp instrument recycling management system based on a module anti-permeation structure according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0145] The first unit is used to construct a medical sharp instrument feature recognition model, collect visible light images, infrared thermal imaging images, and ultrasonic scanning images corresponding to field medical sharp instruments and combine them into medical sharp instrument images, perform a multi-modal image fusion algorithm based on deep hashing on the medical sharp instrument images to obtain fused images, extract the shape, material, texture, and internal structure corresponding to the field medical sharp instruments based on the fused images and combine them to generate medical sharp instrument features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharp instrument features through a combined mechanism of channel attention and spatial attention, enhance the recalibrated medical sharp instrument features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharp instrument features through an improved adversarial training method to obtain independent features, evaluate the importance of the independent features, and establish a feature selection strategy to obtain the key features of the medical sharp instrument;

[0146] The second unit is used to perform based on the key features of the medical sharp instrument, accurately locate the medical sharp instrument through an improved dense prediction network to obtain location features, construct a refined mask for the medical sharp instrument through a segmentation algorithm guided by hierarchical attention to obtain segmentation features, analyze the component composition relationship of the medical sharp instrument using a graph convolutional network to obtain semantic features, fuse the location features, segmentation features, and semantic features and input them into a causal inference network, establish a dependency relationship between features through an improved probability graph model, construct a multi-level classifier based on the dependency relationship, and classify the medical sharp instrument in combination with an ensemble learning strategy to obtain the category information and danger level information corresponding to the medical sharp instrument;

[0147] A third unit is used to construct a classification scheduling system based on the category information and the risk level information. It models the dynamic state representation through an improved spatio-temporal graph neural network and extracts the state representations at different spatio-temporal scales. The state representations are added to a time series prediction model to determine the state evolution trend. An initial scheduling strategy is generated by combining a hierarchical reinforcement learning algorithm. The initial scheduling strategy is optimized through an improved policy gradient algorithm in a pre-set two-stage optimization network. Dynamic allocation is performed by combining an improved auction mechanism to obtain a dynamic allocation result. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeation structure. The dynamic allocation result is adjusted through an improved optimization algorithm, and the adjusted dynamic allocation result is added to a path planning module. A set of transportation paths is generated by combining an improved ant colony algorithm and a heuristic search strategy. The optimal recovery route is obtained by solving through a Pareto optimization algorithm.

[0148] In a third aspect of the embodiments of the present invention,

[0149] a kind of electronic device is provided, including:

[0150] a processor;

[0151] a memory for storing instructions executable by the processor;

[0152] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0153] In a fourth aspect of the embodiments of the present invention,

[0154] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0155] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0156] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for the recovery and management of medical sharps in the wild based on a module anti-penetration structure, characterized in that, Including: Construct a medical sharp object feature recognition model. Collect visible light images, infrared thermal imaging images, and ultrasonic scanning images corresponding to field medical sharp objects and combine them into medical sharp object images. Execute a multi-modal image fusion algorithm based on deep hashing on the medical sharp object images to obtain fused images. Extract the shape, material, texture, and internal structure corresponding to the field medical sharp objects from the fused images and combine them to generate medical sharp object features. Construct a multi-scale feature pyramid through a hierarchical attention mechanism, and adaptively recalibrate the medical sharp object features by combining a combination mechanism of channel attention and spatial attention. Enhance the recalibrated medical sharp object features through a deformable convolutional network and a residual connection structure. Use an improved adversarial training method to decouple the enhanced medical sharp object features to obtain independent features. Evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of the medical sharp objects; Based on the key features, accurately locate the medical sharp objects through an improved dense prediction network to obtain location features. Use a hierarchical attention-guided segmentation algorithm to construct a refined mask for the medical sharp objects to obtain segmentation features. Analyze the component composition relationship of the medical sharp objects using a graph convolutional network to obtain semantic features. After fusing the location features, segmentation features, and semantic features, input them into a causal inference network. Establish a dependency relationship between features through an improved probabilistic graph model. Based on the dependency relationship, construct a multi-level classifier, and classify the medical sharp objects by combining an ensemble learning strategy to obtain the category information and danger level information corresponding to the medical sharp objects; Construct a classification scheduling system based on the category information and the danger level information. Model the dynamic state representation through an improved spatio-temporal graph neural network and extract the state representations at different spatio-temporal scales. Add the state representations to a time series prediction model and determine the state evolution trend. Combine a hierarchical reinforcement learning algorithm to generate an initial scheduling strategy. Optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network. Combine an improved auction mechanism for dynamic allocation to obtain a dynamic allocation result and construct a safety constraint model based on the geometric parameters and material parameters of the anti-penetration structure. Adjust the dynamic allocation result through an improved optimization algorithm and add the adjusted dynamic allocation result to a path planning module. Combine an improved ant colony algorithm and a heuristic search strategy to generate a set of transportation paths, and solve through a Pareto optimization algorithm to obtain the optimal recovery route.

2. The method according to claim 1, characterized in that, Build a medical sharp feature recognition model, collect visible light images, infrared thermal imaging images, and ultrasonic scan images corresponding to field medical sharps and combine them into medical sharp images, perform a multi-modal image fusion algorithm based on deep hashing on the medical sharp images to obtain a fused image, extract the shape, material, texture, and internal structure corresponding to the field medical sharps from the fused image and combine them to generate medical sharp features, construct a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharp features by combining the combined mechanism of channel attention and spatial attention, enhance the recalibrated medical sharp features through a deformable convolutional network and a residual connection structure, use an improved adversarial training method to decouple the enhanced medical sharp features to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of the medical sharp, including: Collect visible light images, infrared thermal imaging images, and ultrasonic scan images of medical sharps, respectively construct feature extraction branches for the visible light images, infrared thermal imaging images, and ultrasonic scan images, the feature extraction branches include multiple convolutional layers, and each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function, and map the extracted features to feature vectors through a fully connected layer; Build a hash coding layer, map the feature vectors to binary hash codes through a fully connected layer and a hyperbolic tangent activation function, calculate the Hamming distance between the binary hash codes corresponding to the visible light images, infrared thermal imaging images, and ultrasonic scan images, screen feature-similar images based on a preset distance threshold, calculate the fusion weight coefficients of the feature-similar images using the softmax function, and linearly combine the weighted features to obtain a fused feature; Build a multi-layer feature pyramid network, adjust the number of channels of the fused feature through convolution, use max pooling operations to downsample layer by layer to obtain feature maps of different scales, calculate the channel attention weights and spatial attention weights for the feature maps, the channel attention weights are calculated through global average pooling and multi-layer fully connected, and the spatial attention weights are calculated through convolution operations and a sigmoid activation function, and multiply the channel attention weights and spatial attention weights to obtain a comprehensive attention map; Build an offset prediction branch containing multiple convolutional layers, predict offset values for each position of the fused feature, limit the offset values within a preset range through a hyperbolic tangent function, adjust the sampling positions of the standard convolution according to the offset values, calculate the feature values at the sampling points using bilinear interpolation and perform convolution operations to obtain enhanced features, and perform a residual connection between the enhanced features and the input features to obtain optimized features; Construct a generator network and a discriminator network. The generator network adopts an encoder-decoder structure. The encoder contains multiple downsampling blocks, each of which includes convolution, instance normalization, and rectified linear unit activation. The decoder contains multiple upsampling blocks, each of which includes transposed convolution, instance normalization, and rectified linear unit activation. The discriminator network contains multiple convolutional layers and adopts a leaky rectified linear unit activation function. Through the generator network and the discriminator network, adversarial training is performed on the optimized features to obtain independent features; Calculate the mutual information value between the independent features and the medical sharp object category labels, perform normalization processing on the independent features, statistically analyze the probability distributions of the independent features and the category labels, calculate the mutual information value based on the probability distributions, and select the features with the highest mutual information value as the key features.

3. The method according to claim 2, wherein The Hamming distance between the computed binary hash codes is shown by the following formula: where D ( H 1, H 2) represents the Hamming distance between the first hash code H 1 and the second hash code H 2, n represents the number of sub - blocks into which the hash code is divided, w (i) represents the local sensitivity weight of the i th sub - block, m represents the number of bits of the hash code in each sub - block, C i , j represents the correlation between the i th hash code position and the j th hash code position, represents the value of the H th binary bit of the first hash code i 1 in the j th sub - block, represents the value of the H th binary bit of the second hash code i 2 in the j th sub - block.​ 4. The method according to claim 1, characterized in that, Based on the key features, accurately locate the medical sharp objects through an improved dense prediction network to obtain location features, construct a refined mask for the medical sharp objects using a hierarchical attention-guided segmentation algorithm to obtain segmentation features, analyze the component composition relationship of the medical sharp objects using a graph convolutional network to obtain semantic features, fuse the location features, segmentation features, and semantic features and then input them into a causal inference network, establish a dependency relationship between the features through an improved probabilistic graph model, construct a multi-level classifier based on the dependency relationship, and classify the medical sharp objects in combination with an ensemble learning strategy to obtain the corresponding category information and danger level information of the medical sharp objects, including: Construct an improved dense prediction network, and locate the medical sharp objects based on the key features. The dense prediction network includes a main branch and an auxiliary branch. The main branch contains multiple dense connection blocks, each of which contains multiple convolutional units, and each convolutional unit is composed of a batch normalization layer, a rectified linear unit activation function, and a convolutional layer. The auxiliary branch contains a spatial pyramid pooling module and a feature pyramid attention module. Output the position coordinates and confidence scores of the medical sharp objects through the dense prediction network to generate location features; Determine the region of interest based on the position coordinates of the location features, construct a hierarchical attention-guided segmentation network within the region of interest. The segmentation network includes an encoder and a decoder. The encoder extracts multi-level features using an improved residual network structure according to the confidence scores of the location features. The decoder performs feature fusion through skip connections according to the multi-level features, and introduces a channel attention module and a spatial attention module to enhance the features during the feature fusion process. The weights of the channel attention module and the spatial attention module are adaptively adjusted according to the confidence scores of the location features. At the same time, an edge-aware module is introduced to optimize the edge details of the enhanced features to generate segmentation features; Input the segmentation features into a graph convolutional network to analyze the component composition relationship of medical sharps. Based on the edge information of the segmentation features, construct a graph structure, where the nodes in the graph structure represent different component regions, and the edges represent the connection relationships between adjacent components. The graph convolutional network includes multiple graph convolutional layers, and the feature transformation parameters of each graph convolutional layer are dynamically adjusted according to the regional characteristics of the segmentation features, and semantic features are generated through feature aggregation; According to the spatial distribution of the positioning features, the regional segmentation results of the segmentation features, and the component relationships of the semantic features, adopt an attention-guided adaptive fusion strategy to calculate feature weights, and input the weighted features into a causal inference network. The temporal flow of the causal inference network models the temporal dependence relationship based on the temporal evolution law of the features, and the spatial flow models the spatial dependence relationship based on the regional distribution characteristics of the features. Based on the dependence relationship output by the causal inference network, construct a probability graph model, map the dependence relationship to the node connection relationship of a directed acyclic graph, and optimize the graph structure and determine the conditional probability distribution in combination with prior knowledge; Based on the positioning features, the segmentation features, and the semantic features, extract the overall contour features, surface color distribution features, and texture features of the medical sharp respectively, combine them to obtain the global features corresponding to the medical sharp, and extract the edge detail features, handheld part features, and functional features of the medical sharp to obtain the local features corresponding to the medical sharp; Construct a multi-level classifier according to the output of the probability graph model, including a coarse classifier for category division and a fine classifier for fine classification. Among them, the coarse classifier uses a random forest structure to classify the global features, and the fine classifier uses a gradient boosting tree structure to perform fine classification on the local features. The classification results are fused by voting. The multi-level classifier is used as a base classifier, and an ensemble learning strategy is used to construct a classifier set, and the model parameters are optimized through cross-validation to output the category information and risk level information of the medical sharp; 5. The method according to claim 4, wherein The coarse classifier uses a random forest structure to classify the global features, and the fine classifier uses a gradient boosting tree structure to perform fine classification on the local features. The fusion of classification results by voting includes: Obtain the global features corresponding to the medical sharp, perform feature normalization processing, construct an improved random forest structure containing multi-level decision units as the coarse classifier, set a feature selection module and a classification module in the improved random forest structure, perform adaptive feature sampling based on the feature importance score through the feature selection module, and use the classification module to construct a decision tree based on information gain and apply regularization constraints to output the basic category of the medical sharp; Obtain the local features corresponding to the medical sharp, establish a cascaded gradient boosting tree structure as the fine classifier, divide the cascaded gradient boosting tree structure into a feature extraction layer and a classification layer, perform attention weighting on the local features by using the feature extraction layer, perform residual learning through the classification layer and iteratively optimize the tree structure parameters, and implement sample balancing and feature interaction operations to output the fine category of the medical sharp; An adaptive multi-modal fusion mechanism is established, a dynamic weight calculation module and a decision fusion module are configured. The dynamic weight calculation module is used to adjust the weight coefficient according to the classification effect, and the decision fusion module performs multi-level voting and initiates expert knowledge base arbitration when the confidence levels are inconsistent, and the final classification result is output by combining the usage scenario and safety risk assessment.

6. The method according to claim 1, wherein Based on the category information and the hazard level information, a classification scheduling system is constructed. The dynamic state representation is modeled by an improved spatio-temporal graph neural network and the state representations at different spatio-temporal scales are extracted. The state representations are added to a time series prediction model to determine the state evolution trend. An initial scheduling strategy is generated by combining a hierarchical reinforcement learning algorithm. The initial scheduling strategy is optimized by an improved policy gradient algorithm in a pre-set two-stage optimization network. Dynamic allocation is performed by combining an improved auction mechanism to obtain a dynamic allocation result. A safety constraint model is constructed based on the geometric parameters and material parameters of the anti-permeation structure. The dynamic allocation result is adjusted by an improved optimization algorithm and the adjusted dynamic allocation result is added to a path planning module. A set of transportation paths is generated by combining an improved ant colony algorithm and a heuristic search strategy. The optimal recovery route is obtained by solving with a Pareto optimization algorithm, including: A spatio-temporal relationship graph of medical sharps is constructed. The medical sharps are used as nodes in the spatio-temporal relationship graph. The nodes contain the position information, category information and hazard level information of the medical sharps. The spatial distribution characteristics of the medical sharps in different regions are analyzed by setting up multiple layers of spatial convolutional layers. The time distribution law of the medical sharps is analyzed by setting up a time convolutional layer. The spatial distribution characteristics and the time distribution law are input into a feature fusion module to generate a state representation of the medical sharps; The state representation is input into a time series prediction model. The time series prediction model includes a time series encoder and a time series decoder. The time series encoder extracts periodic features and sudden features from the historical distribution data and historical recovery data of medical sharps according to the time order. The time series decoder predicts the distribution trend of medical sharps based on the periodic features and the sudden features. The historical distribution data and the historical recovery data are weighted by an attention mechanism to generate a state evolution trend; A hierarchical reinforcement learning network is constructed based on the state evolution trend. The hierarchical reinforcement learning network includes a high-level policy network and a low-level execution network. The high-level policy network divides the recovery tasks of medical sharps into priorities according to regions and types based on the state evolution trend to obtain a set of sub-tasks. The low-level execution network determines the recovery time and recovery route based on the set of sub-tasks to obtain an initial scheduling strategy. The initial scheduling strategy is stored in an experience replay pool and scheduling experiences are selected by priority sampling; Input the initial scheduling policy into the two-stage optimization network. In the first stage of the two-stage optimization network, the value of the scheduling action is evaluated through the policy gradient algorithm, and the entropy regularization term is introduced to adjust the policy parameters. In the second stage of the two-stage optimization network, resource allocation for medical sharps is based on the auction mechanism, and a dynamic allocation result is generated according to the bidding results of the collection points and the constraint conditions. Construct a security constraint model based on the dynamic allocation result. The security constraint model selects anti-permeation containers according to the physical characteristics and hazard levels of medical sharps, monitors the loading conditions of the anti-permeation containers in real time, automatically adjusts the transportation plan when security hazards are detected, and inputs the transportation plan into the path planning module. The path planning module uses the ant colony algorithm to construct a path graph between collection points, conducts path search based on transportation distance, time window, and security factors, and optimizes the path sequence through pheromone update and local search to generate an optimal path set that meets multi-objective constraints.

7. The method according to claim 6, characterized in that In the second stage of the two-stage optimization network, resource allocation for medical sharps is based on the auction mechanism, and the generation of a dynamic allocation result according to the bidding results of the collection points and the constraint conditions includes: Construct bidding entities. Set the medical sharp collection points as the buyer entities and the medical sharps to be recycled as the seller entities. The medical sharp collection points have preset bidding budgets and processing capacity constraints. The processing capacity constraints include the number of medical sharps that can be processed per unit time and the processing capacity limits for different types of medical sharps. Set the reserve price based on the category information, hazard level, and urgency of the medical sharps to be recycled. Adopt a multi-round ascending bidding method to construct the bidding rules. The medical sharp collection points submit bids that meet the maximum bid limit, budget limit, and processing capacity requirements in each round of bidding. When the medical sharp collection points submit new bids, start the price update mechanism, and the remaining medical sharp collection points follow up with bids within the specified time, and promote the bidding process based on the minimum price increment limit. Construct a bidding strategy optimization mechanism based on historical bidding data and the current market situation. Predict the optimal bid range by analyzing historical transaction prices, competitors' bidding patterns, and market supply and demand relationships. Set dynamic allocation constraint conditions including time window constraints, path constraints, and capacity constraints. The time window constraint is used to ensure the time limit requirement for the recycling of medical sharps, the path constraint is used to ensure the feasibility of the transportation route, and the capacity constraint is used to limit the maximum processing volume of the medical sharp collection points. Construct a multi-objective evaluation system including the allocation efficiency dimension, cost control dimension, and risk management dimension. The allocation efficiency dimension includes the timeliness index and balance index of medical sharp recycling. The cost control dimension includes the transportation cost index and processing cost index. The risk management dimension includes the security hazard index and emergency response capacity index. Generate a preliminary allocation plan based on the bidding results and the dynamic allocation constraint conditions, and conduct a feasibility verification on the preliminary allocation plan. When constraint conflicts occur, generate a dynamic allocation result that meets the actual needs by adjusting the allocation quantity or replacing the medical sharp collection points.

8. A field medical sharp waste recycling management system based on a module anti-permeation structure, for implementing the method according to any one of the preceding claims 1-7, characterized in that, Including: The first unit is used to build a medical sharp object feature recognition model, collect visible light images, infrared thermal imaging images, and ultrasonic scanning images corresponding to field medical sharp objects and combine them into medical sharp object images, perform a multi-modal image fusion algorithm based on deep hashing on the medical sharp object images to obtain fused images, extract the shape, material, texture, and internal structure corresponding to the field medical sharp objects from the fused images and combine them to generate medical sharp object features, build a multi-scale feature pyramid through a hierarchical attention mechanism, adaptively recalibrate the medical sharp object features through a combined mechanism of channel attention and spatial attention, enhance the recalibrated medical sharp object features through a deformable convolutional network and a residual connection structure, decouple the enhanced medical sharp object features through an improved adversarial training method to obtain independent features, evaluate the importance of the independent features and establish a feature selection strategy to obtain the key features of the medical sharp objects; The second unit is used to execute based on the key features of the medical sharp objects. Precise positioning of the medical sharp objects is obtained through an improved dense prediction network to obtain positioning features, a refined mask for the medical sharp objects is constructed through a segmentation algorithm guided by hierarchical attention to obtain segmentation features, the component composition relationship of the medical sharp objects is analyzed through a graph convolutional network to obtain semantic features, the positioning features, segmentation features, and semantic features are fused and then input into a causal inference network, a dependence relationship between features is established through an improved probability graph model, a multi-level classifier is constructed based on the dependence relationship, and the medical sharp objects are classified by combining an ensemble learning strategy to obtain the category information and danger level information corresponding to the medical sharp objects; The third unit is used to build a classification scheduling system based on the category information and the danger level information, model dynamic state representations through an improved spatio-temporal graph neural network and extract state representations at different spatio-temporal scales, add the state representations to a time series prediction model and determine the state evolution trend, generate an initial scheduling strategy by combining a hierarchical reinforcement learning algorithm, optimize the initial scheduling strategy through an improved policy gradient algorithm in a pre-set two-stage optimization network, perform dynamic allocation by combining an improved auction mechanism to obtain a dynamic allocation result and build a safety constraint model based on the geometric parameters and material parameters of the anti-penetration structure, adjust the dynamic allocation result through an improved optimization algorithm and add the adjusted dynamic allocation result to a path planning module, generate a set of transportation paths by combining an improved ant colony algorithm and a heuristic search strategy, and solve through a Pareto optimization algorithm to obtain the optimal recovery route.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images