1750 results about "Global information" patented technology

Provided is a method for training a scene recognition model. The scene recognition model includes a core feature extraction layer, a global information feature extraction layer, an LCS module of at least one level with an attention mechanism, and a fully-connected decision layer. The method includes: acquiring parameters of the core feature extraction layer and the global information feature extraction layer by training based on a first scene label of a sample image and a standard cross-entropy loss; training a weight parameter of the LCS module of each level, based on a loss value acquired by performing a pixel-by-pixel calculation on a feature map output from the LCS module of each level and the first scene label of the sample image; and acquiring a parameter of the fully-connected decision layer by training based on the first scene label of the sample image and the standard cross-entropy loss.

The invention provides a remote sensing image super-resolution method based on a diffusion model and a multi-modal large language model, and relates to the technical field of image processing. According to the method, a pre-trained stable diffusion model is adopted, the model is trained on a large-scale data set, the stable and detailed super-resolution effect can be achieved, and the convergence speed is higher. According to the method, a multi-mode large language model is used for enhancing the semantic comprehension ability of the model, and detailed content description and texture description are provided. The invention also introduces an additional classifier which provides global information and a simple signal about the LR. The prior information is adaptively integrated into a diffusion model through a cross attention module. And finally, through combination of the strong generation capability of the stable diffusion model and the semantic understanding capability of the MLLM, the method provided by the invention is excellent in the aspect of remote sensing image reconstruction, and excellent performance is realized in the aspect of quality of a plurality of images on different data sets.

According to the medical image segmentation method and device based on spatial perception and frequency domain information, the precision and robustness of medical image segmentation are effectively improved by combining frequency domain information guidance and a multi-head state spatial perception technology. Frequency domain transformation is performed on a medical image, high-frequency and low-frequency components of the image are separated by using a multi-scale decomposition technology, and low-frequency features are extracted to obtain global information. By introducing a learnable noise filtering mechanism, noise and irrelevant background information in a frequency domain are suppressed, so that the model can be focused on a lesion area more accurately. A multi-head perception visual state space module is designed on a bottleneck layer, lesion features of different scales are captured through a multi-scale adaptive feature fusion mechanism, and the capability of segmenting small-size lesions and complex structures is enhanced. A context focusing attention mechanism is introduced into jump connection, fusion of global information and local details is further enhanced, and the accuracy of a segmentation result is ensured; and finally, recovering a high-resolution segmented image through a decoder.

The invention relates to the technical field of medical image segmentation, in particular to a few-sample medical image segmentation method based on edge perception multi-prototype learning, and the method comprises the steps: inputting support and query images into a feature encoder, and extracting support and query feature maps of different sizes; inputting into a local attention fusion prototype generator to generate a support foreground prototype; processing the support mask through dynamic corrosion operation to generate an inner boundary prototype; generating a multi-foreground local prototype through a multi-layer perceptron; local and global information is optimized through multi-scale feature extraction, and a multi-scale prototype is obtained; fusing to obtain a multi-prototype foreground prototype; dynamic calculation weighting is carried out on the multi-prototype foreground prototype by using a double-stage prototype optimization network, and automatic calibration is carried out; then prediction is carried out through a prototype prediction module, and finally collaborative optimization is carried out through a loss calculation module; the method can effectively solve the problem of edge detail loss involved in the background technology.

The invention discloses a de-noising diffusion model driven texture enhanced infrared and visible light image fusion method and system, and relates to the technical field of image processing and computer vision. The method comprises the following steps: performing multi-scale space-time modeling on infrared and visible light images; enhancing high-frequency texture features of the infrared image through an edge detection convolutional network; based on a bidirectional multi-scale convolution module and a bidirectional attention fusion module, capturing a remote dependency relationship and encoding multi-scale information to ensure full utilization of global information and accurate capture of local details; and optimizing the structural fidelity and color intensity distribution of the fusion result based on an adaptive structural similarity loss function and multi-channel intensity and texture loss so as to improve the quality of the fused image. According to the method, the process of diffusion feature extraction is enhanced by fusing the de-noising diffusion model and the texture enhancement strategy. Meanwhile, a bidirectional multi-scale convolution technology and an attention mechanism are combined, so that comprehensive utilization of global information and accurate mastering of local details are ensured.

The invention relates to the technical field of biomedical engineering, in particular to a multi-mode brain anomaly detection method and system based on machine learning. The method comprises the following steps: acquiring brain medical image data of different modalities, and realizing spatial registration and alignment through a multi-modal registration algorithm based on mutual information; a multi-branch feature extraction model including a convolutional neural network, a converter and a state space model is utilized to perform feature embedding on the original image of each modal; performing frequency decoupling on the features of each mode through adaptive approximate wavelet transform, and decomposing the features into high-frequency detail information and low-frequency global information; a frequency band fusion strategy based on an attention mechanism is implemented on high and low frequency features of different modal images, and fused frequency sub-band features are input into a space-frequency Mama module. Through the adaptive frequency domain decomposition and cross-modal fusion mechanism, the multi-modal brain image information is effectively integrated, and the accuracy and robustness of brain anomaly detection are remarkably improved.

The invention belongs to the field of outdoor AGV obstacle tracking, and particularly relates to an AGV obstacle tracking method based on multi-modal data and interactive Kalman filtering. Comprising the following steps: calibrating a monocular camera and a laser radar, and establishing point cloud data and image data with synchronous timestamps; performing preprocessing and target detection operation on the point cloud data and the target detection frame to obtain preprocessed point cloud data and a generated target detection frame so as to obtain obstacle category information; acquiring an obstacle observation position according to the preprocessed point cloud data and the target detection frame; meanwhile, an improved SORT multi-target tracking algorithm is adopted, and a historical track of the obstacle is established; through a plurality of Kalman filtering fusion methods, a trajectory prediction result is obtained through prediction; finally, the AGV communicates with a dispatching center, global information is obtained, and the track trend is corrected in a targeted mode in combination with obstacle category information; and the historical track of the obstacle and the corrected prediction track are sent to the AGV for obstacle avoidance path planning.

The invention discloses an improved YOLOv11 power transmission line defect detection method, which introduces an attention mechanism to optimize the feature extraction capability, improves the performance of a model in a complex scene, remarkably enhances the understanding of the model for local and global information, and improves the target detection precision. Spatial offset is carried out on a feature map in a detection head by adopting feature extraction, feature association of different scales is enhanced, the detection precision of a small target is improved, redundant calculation is reduced, the calculation efficiency and performance of a model are further optimized, diversified requirements in a complex environment are met, and the method is suitable for popularization and application. The feature learning ability of a self-integrated attention head optimization target detection model is introduced, and the detection precision and robustness are improved under the scenes of small object detection and complex backgrounds. The method effectively improves the detection precision and training efficiency, can achieve excellent performance in a high-precision requirement and a real-time task, and meets the requirements of safety risk management and control in the power industry.

The invention belongs to the field of multi-modal data processing, and provides a multi-modal mixed expert psychological health assessment system, method, medium and equipment in order to solve the problem of low feature fusion efficiency caused by insufficient dynamic interaction between modals of an existing mixed expert model in a multi-modal scene. The multi-modal mixed expert psychological health assessment system comprises a multi-modal data acquisition module, and an adaptive grouping module. The cross attention fusion module is used for extracting features of corresponding modal data in the basic group and the auxiliary group, performing intra-group cross attention fusion on the features of each group, and generating intra-group features; performing global information fusion on all the intra-group features through a cross-group cross attention mechanism to obtain cross-group features; and after the cross-group features, the intra-group features and the original features of the multi-modal data are spliced, a dynamic gating weight is obtained through a gating network. The method can provide more accurate guidance for psychologists.

The invention belongs to the technical field of target detection, and particularly relates to a student classroom behavior detection method based on deep learning. The method comprises the following steps: training an improved YOLOv11 model, and pruning the trained improved YOLOv11 model to obtain a student classroom detection model; acquiring a to-be-detected student classroom video and inputting the to-be-detected student classroom video into the student classroom detection model for processing to obtain a student classroom behavior detection result; the improved DeepSORT algorithm is used for processing the student classroom behavior detection result, tracking of the student classroom behaviors is achieved, the global information capturing capacity of the network is enhanced, the performance of the network under the shielding and multi-scale change scene is improved, and therefore the robustness and detection precision of the network are improved, and the method has good application prospects.

The invention belongs to the field of target detection, and particularly relates to an infrared weak and small target detection method based on a double-branch attention mechanism. The method comprises the following specific steps: step 1, preparing data sets: performing data enhancement on a first infrared data set and a second infrared data set for pre-training, and performing model fine adjustment on a third infrared data set; 2, constructing a network model, wherein the whole model is composed of a target analysis path, a splicing path and a target reconstruction path; according to the method, a global semantic enhancement large-kernel attention module is introduced into a coding module of a target analysis path, global information is obtained and utilized from four angles of a channel, a transverse direction, a longitudinal direction and a receptive field, and a double-branch single-head attention module is introduced into a splicing path to obtain association degrees among different scales; the purpose of eliminating the semantic difference between the target analysis path and the target reconstruction path is achieved, the difference between the weak and small target and the background is expanded, and the detection precision of the infrared weak and small target is improved.

The invention relates to a micro-expression recognition method based on a cross-source double-branch dynamic space-time diagram convolutional network model, and belongs to the technical field of deep learning and pattern recognition. According to the method, a cross-source double-branch twin space-time diagram structure network is provided to mine subtle motion features of a facial structure when expressions change, and domain invariant features are learned through a twin structure. According to the design, facial global information is modeled through a global information and dynamic spatio-temporal feature extraction network, fine motion information of a facial key structure is extracted through an attention enhancement twin spatio-temporal diagram fusion network, facial structure association when expressions occur is fully modeled, and domain adaptation is performed by constructing cross-domain joint loss.

The invention relates to the technical field of industrial defect detection, and discloses an overflowing surface defect detection and identification method and system of a multi-modal feature and prompt mechanism, and the method comprises the steps: collecting image data and point cloud data of the surface of a water turbine runner overflowing surface, and converting the point cloud data into a surface defect two-dimensional image; obtaining a high-resolution defect image based on an image super-resolution reconstruction method of a deep recursive network; inputting a defect image into the generative adversarial network, judging a data source, generating a high-quality defect image through an objective function optimization generative model, and expanding a data set for a defect identification detection model to use; and inputting the point cloud data into the PDE-Net to construct a dynamic adjacency graph, and mapping the point cloud data to a specific defect category after graph convolution and multi-layer feature extraction. The method has the beneficial effects that a mixed attention mechanism is introduced in a multi-modal feature extraction stage, global information and local details in a defect image are fully extracted, and the detection capability of complex defects is improved.

The invention belongs to the technical field of medical image processing, and particularly relates to a semi-supervised cardiac image segmentation method based on a Mamb-Transform double structure and contrast learning. A segmentation model used in the method comprises an MTSeg encoder and a VNet decoder, an input image enters the MTSeg encoder after being subjected to linear layer and position embedding operation, the MTSeg encoder comprises a plurality of MTSeg encoding modules, the VNet decoder comprises a plurality of decoding modules, and output features of the MTSeg encoding modules are in jump connection with the corresponding decoding modules after being subjected to projection operation; performing normalization operation on the output feature vector of the VNet decoder to obtain a segmentation result; the MTSeg coding module comprises an improved Mama branch and a Transform branch, and output feature vectors of the two branches are fused through a cross attention layer to obtain an output feature vector of the MTSeg coding module; the improved Mama branch comprises an MS module and a DRFB module which are connected in series, the MS module is used for enhancing the feature extraction and calculation efficiency of the model, and the DRFB module is used for enhancing the expression ability and training stability of the model. The method achieves the effective integration of a long-range dependency relationship, global information, multi-scale information and local information, improves the feature extraction capability of the model, and obtains a finer and more accurate segmentation result.

The invention relates to a multi-modal data fusion method and device, computer equipment and a storage medium. The method comprises the following steps: carrying out high-order feature extraction on multi-modal data to be fused by utilizing a special modal encoder to obtain high-order embedding features of the multi-modal data; the high-order embedded features are decoupled in combination with expert knowledge constraints and a multi-head attention mechanism, and high-frequency features representing local information and low-frequency features representing global information are obtained respectively; and aligning the low-frequency features by using a cosine alignment algorithm, performing orthogonal projection on the high-frequency features by using an orthogonal projection technology, and performing multi-level fusion on the aligned low-frequency features and the orthogonally projected high-frequency features through an adaptive deep fusion network to obtain fused multi-modal joint features. According to the method, the problems of heterogeneity and isomerism of the data are solved, effective fusion of different modal data is ensured, and the accuracy and individuation capability of disease risk early warning can be remarkably improved.

The invention belongs to the technical field of natural language processing and multi-modal information extraction, and particularly relates to a multi-modal named entity recognition method based on semantic alignment and cross-modal graph fusion, which comprises the following steps: S1, acquiring a data sample containing a text sequence and image content; s2, encoding the text and the image into vectors respectively; s3, similarity is calculated through a trainable bilinear function, and optimization is carried out through loss comparison; s4, cross-modal attention is used to enhance association information between modals; s5, determining the proportion of reserved image information through a modal matching module; s6, introducing a gating mechanism to dynamically fuse visual and text features; s7, realizing local and global information complementation by a cross-modal graph fusion model; and S8, inputting the fused representation into the CRF layer to predict the entity type. According to the method, fine semantic alignment can be realized in a weak image-text correlation context, and balance between local entity recognition and global semantic understanding can be achieved.

The invention discloses an aerospace network hierarchical cooperation routing optimization method and system, and relates to the field of aerospace information network communication, and the method comprises the steps: constructing a system model comprising an unmanned plane, satellite nodes and a communication link, defining link attributes, building a routing optimization objective function, introducing a reputation value-based multi-agent cooperation mechanism, and carrying out the routing optimization of the unmanned plane, the satellite nodes and the communication link. A hierarchical reinforcement learning framework is designed, inter-cluster decision and intra-cluster decision are divided, an upper layer selects an inter-cluster path according to global information, a lower layer plans an unmanned aerial vehicle path and defines a state, an action space and a reward function according to local information, an intelligent agent interacts with an environment to collect empirical data, and optimal routing communication is selected in combination with a real-time reputation value after training. Adjustment can be carried out according to network changes; according to the invention, the cooperative reliability is improved through a reputation value mechanism, the decision complexity is reduced through hierarchical learning, the routing performance is optimized, the network robustness, expandability and adaptability are enhanced, and the cooperative communication efficiency of the unmanned aerial vehicle and the satellite and the system stability are improved.

The invention discloses an aircraft 4D trajectory prediction method and system based on double-attention space-time diagram convolution, and belongs to the technical field of intelligent time sequence prediction. According to the method, the self-attention mechanism is utilized to reconstruct the predefined adjacent matrix, and optimization is continuously carried out along with network training so as to improve generalization of the network. The problem that the distance-based predefined adjacency matrix cannot fully reflect the potential relationship between the aircrafts in the dynamic scene is effectively solved. On the basis of a learnable adjacency matrix based on a self-attention mechanism, the correlation between nodes is further aggregated by using the graph attention network, and additional feature extraction is performed on the space-time diagram based on the powerful global information aggregation capability of the graph attention network, so that the model can capture more potential relationships between the nodes, and the accuracy of feature extraction is improved. And the problem that the correlation between the nodes cannot be effectively aggregated by the traditional graph convolution operation is improved. And the prediction accuracy of the future trajectory of the aircraft is further improved.

The invention relates to a multi-modal understanding optimization method based on fine-grained feature extraction and global information integration. The method comprises the following steps: segmenting an input image into a plurality of local image blocks, extracting local fine-grained visual features and global visual features, and interacting the local fine-grained visual features and the global visual features based on an attention mechanism to obtain local context features; secondly, performing feature fusion on the local fine-grained visual features to generate fused visual features; mapping the fused visual features to a semantic space which is the same as the text features of the large language model to obtain projected visual features, and dynamically screening through attention weight based on the text features to obtain key visual features; and fusing the key visual features with the text features to generate joint feature representation, and inputting the joint feature representation into a large language model to generate a semantic analysis result. According to the method, global information and dynamic feature selection are introduced, so that the ability of the model to understand multi-modal content in a complex scene can be improved, and the model calculation overhead is reduced.

In order to solve the problems that day lily plants are densely shielded and categories are difficult to distinguish, the invention provides a field dense day lily pixel classification and picking information acquisition method. The method comprises the steps of designing a double-branch decoder fusing local and global information, constructing a progressive stepped feature mining module (PLFM) and a multi-scale hierarchical category enhancement module (PCEN), optimizing hyper-parameters through a PALA algorithm, and constructing an adaptive double-branch loss enhancement and category enhancement network (DLCE-Net) to improve semantic segmentation performance. And designing a picking point positioning algorithm (MW-GCA) of Mini-Window guided angular point analysis based on a segmentation result, screening angular point coordinates through a U-shaped window, obtaining an attitude line segment, and realizing picking point pixel coordinate and angle estimation. The method can reduce the labor loss and accelerate the engineering landing of the automatic crop picking technology.

The invention discloses an MRI (Magnetic Resonance Imaging) image segmentation method based on WKA-Unet. The MRI image segmentation method comprises the following steps: 1) data preprocessing; 2) constructing a WKA-Unet network; and 3) testing. According to the technical scheme, on the basis of a U-Net network structure, an MRI image segmentation model is improved; a traditional Unet network is combined with a self-designed WKA module, multi-scale features are fully captured by using wavelet transform convolution and KAN convolution, and the loss of traditional Unet global information is relieved; an attention mechanism is introduced to pay attention to key parts, and redundant information is removed; according to the technical scheme, the network is trained more flexibly and accurately, a more accurate segmentation result is obtained, the segmentation precision is high, and the model structure is efficient and stable.

The invention discloses a pedestrian and vehicle detection method based on an improved YOLOv10n. The pedestrian and vehicle detection method comprises the following steps: on the basis of an original model YOLOv10n, replacing a first layer and a third layer in Backbone and a seventeenth layer Conv in Neck with a space-to-depth convolution SPDConv; a focused spatial pyramid pooling module ASPPF is constructed, and semantic fusion of cross-level features is optimized; designing a global grouping coordinate attention module GGCA, fusing the GGCA into a Neck layer, generating attention maps by using global information of the feature maps in the spatial dimension, and weighting the input feature maps through the attention maps to enhance the feature expression ability; a WIou loss function is introduced, the bounding box loss of the network is optimized, and the convergence speed and regression precision of the model are improved. And training a model by using a data set which is made by autonomous labeling to obtain a training result, and completing a pedestrian and vehicle detection task by using an optimal training weight. According to the method, a good detection effect is achieved with a small number of model parameters, and the pedestrian and vehicle recognition accuracy is well improved.

The invention discloses an underwater target recognition method based on an improved YOLOv8 algorithm, and the method comprises the steps: inputting an underwater image into an improved YOLOv8n model for underwater target detection, and obtaining an output underwater target recognition result. The method has the advantages that the convolution blocks of the P5 layer of the backbone network and the last layer of the neck network adopt DSConv, so that the network complexity is reduced, and the reasoning speed is increased; a fourth C2f module of the backbone network adopts a C2f DiRMB module in which an inverted residual attention mechanism and dual-channel convolution are introduced, so that the capability of capturing key global information of the network is enhanced, training parameters are reduced, and the understanding of a complex scene is improved; and finally, a small target detection head for improving the small target detection capability is additionally arranged in the head network. According to the underwater target identification method, the mAP (at) is 0.5%, the mAP (at) is 0.5-0.95%, and the accuracy and the recall rate are respectively improved by 0.5%, 0.8%, 0.5% and 1.0%.

The invention discloses a railway turnout defect detection method based on an SCG-DETR detection model. Firstly, a C2f-SMP feature extraction module is designed based on SMPConv and C2f modules, local and global information aggregation of features is optimized through continuous convolution and a gating mechanism, and the receptive field and feature extraction capability is enhanced. Secondly, a CGFM module is provided, and by introducing a CAA attention mechanism, weight distribution of a feature map is optimized, and fine-grained extraction and context information fusion are enhanced. And finally, replacing an original loss function with an Inner-GIoU based on an auxiliary frame, and improving the precision of the model while accelerating the convergence of the model. Experimental results show that the score of mAP50 of the improved model reaches 71.0% and is increased by about 4.6%, the detection speed reaches 80.5 FPS, and the requirement for real-time performance in engineering is met.

The invention discloses an ultrasonic mammary gland image segmentation method based on fine-grained guidance, relates to the technical field of image segmentation, and attempts to use a fine-grained guidance attention decoding diffusion model. The defects of weak association of foreground and background semantics, unbalance of global and local features, poor fine-grained feature capture capability and the like in a traditional segmentation method are overcome. The core design is as follows: a clear de-noised image is obtained by using a de-noising diffusion probability model; based on a context cross-attention decoding diffusion network guided by fine granularity and a contained self-adaptive detail guide attention module, image prior information is learned and fused, and semantic correlation between a foreground and a background is enhanced; a context decoding cross attention layer is added, global information of input features and complex correlation between channels are effectively captured, and the image segmentation precision and efficiency are integrally improved. A contrast experiment result on a related data set shows that compared with a similar popular method, the method disclosed by the invention obtains a better image segmentation effect.

The invention provides a personalized federal learning method and system for a shared model, and the method comprises the steps: enabling a server to store and initialize the shared model, a global model and a global class prototype of a client, and enabling the client to initialize a local learning weight vector; the server sends sharing models of other clients to one client, and the client sets a local sharing model as a global model and trains the global model, and uploads the sharing model and a local class prototype to the server; and the server calculates and obtains a global class prototype and a global model according to all the received shared models and local class prototypes. According to the method, the knowledge sharing problem in federal learning is solved, and the performance of a personalized model is improved; the problem of offset in local model training of the client is solved, and the obtained personalized prototype contains more abundant global information than the personalized prototype; while state dependence and communication overhead are reduced, effective fusion of global knowledge and local characteristics is realized so as to improve robustness and adaptability of the model in a complex scene.

The invention discloses a multi-source tight coupling SLAM method based on global information in a dynamic scene, and belongs to the field of mobile robots. According to the method, efficient positioning and mapping are realized through multi-source sensor data fusion and algorithm optimization. The method comprises the following steps of: firstly, eliminating invalid point cloud and data distortion by combining a binocular camera, a single-line laser radar, an IMU (Inertial Measurement Unit) and odometer data and utilizing point cloud filtering, dynamic target detection and motion distortion correction technologies, and fusing observation data of the camera and the laser radar; secondly, IMU and odometer data are fused through unscented Kalman filtering (UKF), and high-precision initial pose estimation is output; and finally, improving a Gmapping algorithm, optimizing a grid map updating module by adopting a subgraph fusion and breadth-first search strategy, and enhancing the environmental feature retention capability of the SLAM module. In a dynamic and complex scene, the method has better robustness, the constructed map better conforms to a real environment, and meanwhile, the pseudo obstacle characteristics in the map can be effectively reduced.

The invention discloses a solar wind speed prediction method based on a fusion prediction mode and a collaborative attention mechanism. The method comprises the following specific steps: collecting and processing solar wind speed and related data within a certain time; dividing the data into a plurality of patches through a patterning segmentation method, learning a plurality of modes in prediction information, and guiding mode fusion by using historical information to obtain fused prediction mode information; learning historical information and fused prediction mode information by using an attention mechanism to obtain time sequence feature representation, aggregating global information through down-sampling and the attention mechanism, and learning inter-variable relation feature representation of variable dimensions for the global information by using a collaborative attention mechanism; mapping the relationship feature representation among the variables through a flat layer to obtain a solar wind speed prediction result; and training and optimizing the model. According to the method, prediction mode information can be fully mined so as to realize interaction between historical information and prediction information, meanwhile, the relation strength between variables can be captured, and the prediction precision is improved.

The invention discloses a spatial domain identification method based on a multi-view weighted fusion GCN network, and the method comprises the steps: taking the gene expression and spatial position of spatial transcriptome data as input, and combining the spatial information based on different similarity measurement modeling with the gene expression after modeling to construct a multi-view weighted fusion graph convolutional network model; and reconstructing a gene expression matrix by using a decoder, capturing global information of ST data, calculating spatial regularization constraint loss by using similarity information and spatial neighbor information, and completing identification of a spatial transcriptome spatial domain. According to the method provided by the scheme, the model can extract the spatial information from the whole situation and the local situation, so that the spatial information is more fully and comprehensively mined, and accurate spatial domain identification is realized.

The invention discloses a brain MRI image segmentation method based on an RWKV model, and belongs to the technical field of medical image processing and artificial intelligence crossing. Firstly, an RWKV linear self-attention module is introduced into a U-Net network framework, and association between remote pixels is established with relatively low calculation overhead, so that the recognition precision of a brain tumor area is improved, the reasoning time is effectively controlled, and the practicability of a model is enhanced. Secondly, according to the method, multi-scale coding and a feature fusion mechanism are combined, local and global information is extracted in a combined manner, features are mined from different resolution levels, and adaptive fusion is realized, so that the fine-grained segmentation effect is improved. And finally, in order to enhance the generalization ability and lightweight deployment performance of the model, the network structure is further optimized, and the overall computing resource demand is reduced, so that the method can better adapt to cross-patient MRI data in a complex clinical environment, and has good practical value and popularization potential.