49results about How to "Raise attention" patented technology

The invention discloses a self-adaptive multi-level forged voice detection method based on a belief propagation mechanism. The method comprises the following steps: performing feature extraction on voice to be detected to obtain an initial feature vector; reasoning the initial feature vector by using a lightweight model to obtain an initial confidence coefficient representing the voice forgery possibility; based on a low-confidence threshold value and a high-confidence threshold value which are determined through performance optimization, performing multi-level discrimination on the input voice and screening difficult samples; generating expert feature vectors for the difficult samples by adopting a depth feature extraction network regulated and controlled by the initial confidence coefficient; constructing a query vector based on the initial confidence, and fusing expert features through an attention mechanism; and voice authenticity determination is completed according to the fusion features. The initial confidence is used as the cross-stage control quantity, so that the self-adaptive adjustment of the detection process is realized, and the average reasoning delay is reduced while the detection accuracy is ensured.

The application discloses a ramjet lightweight flow field reconstruction method guided by multi-scale feature mapping, and belongs to the technical field of engine combustion monitoring and artificial intelligence application, and the method comprises the following steps: S1, acquiring wall surface pressure time sequence signals and schlieren images, and constructing a combustion flow field reconstruction data set under different inflow conditions; S2, performing data preprocessing on the combustion flow field reconstruction data set to obtain a preprocessed combustion flow field reconstruction data set; S3, inputting the preprocessed combustion flow field reconstruction data set into a space-time multi-scale feature alignment and fusion model to obtain a lightweight flow field reconstruction model; and S4, real-time receiving wall surface pressure time sequence signals and inputting the wall surface pressure time sequence signals into the lightweight flow field reconstruction model to reconstruct a high-time-resolution two-dimensional transient flow field. The method realizes high-precision and high-efficiency reconstruction of extreme combustion non-steady-state multi-physical fields under high-speed flight conditions, and solves the problems of data sparseness and inability to perceive combustion characteristics at future time points in advance in the traditional method.

This invention relates to the field of image generation technology, and more particularly to a method for generating bird's-eye view images of reservoir landscapes based on a Stable Diffusion large model. The method includes the following steps: preprocessing acquired reservoir landscape images and semantically annotating them based on water conservancy terminology to construct an image-text pairing dataset; training a pre-trained Stable Diffusion model, incorporating LoRA adaptation parameters into its UNet attention layer, using the image-text pairing dataset to generate a target feature image; and updating the LoRA adaptation parameters based on the difference between the target feature image and the reservoir landscape image to obtain a LoRA model for the reservoir landscape bird's-eye view image. This invention achieves structurally controllable and semantically consistent reservoir landscape generation by fusing structural reference constraints and semantically weighted alignment and introducing LoRA modulation.

The application discloses a content recommendation method, device and equipment based on co-occurrence matrix optimization. Firstly, the keywords and the word frequency of the keywords in the target text material are used to generate a scoring result vector by using a pre-constructed co-occurrence matrix. In addition, a content transfer matrix is constructed based on the co-occurrence of the content of each pair of target types in each historical text material. A content co-occurrence graph with corresponding content as the vertex and the transfer probability as the edge weight is constructed based on the content involved in the co-occurrence matrix and the content transfer matrix. In the content co-occurrence graph, the semantic weighted score in the scoring result vector is used as the initial weight of the corresponding vertex, and the weight of the vertex in the content co-occurrence graph is iterated by using the content transfer matrix. The content recommendation list of the target text material is generated by combining the weight of each vertex after multiple iterations and the semantic weighted score of each content. The above steps realize effective transmission of important information, improve the accuracy of content recommendation and the utility of recommended content.

The invention belongs to the technical field of milling force prediction, and discloses a milling force prediction system based on a mechanism-data hybrid drive model, which combines an LSTM (Long Short Term Memory) network and a full-connection neural network to construct a milling force prediction model of a deep learning network main body containing residual connection and a Self-Attention mechanism. The model architecture specifically comprises an input layer, an LSTM time sequence feature extraction layer, residual connection, a Self-Attention feature enhancement layer and a full-connection neural network FCN prediction layer. By adopting the milling force prediction system, the attention mechanism effectively enhances the attention of the model on key input by dynamically distributing different weights, so that the prediction precision is improved; the problems of gradient disappearance and gradient explosion are relieved by using residual structure jump connection; the LSTM can effectively capture the long-time dependency relationship in the long-time sequence, and the training efficiency of the model is improved.

This invention provides a method for identifying the authenticity of wheat flour based on the fusion of Raman and near-infrared spectroscopy, belonging to the field of wheat flour authenticity identification technology. The method includes: acquiring Raman and near-infrared spectral data of the wheat flour sample to be tested, and preprocessing them separately; constructing and training a fusion detection model, which includes a data-level fusion module, a feature-level fusion module, and a decision-level fusion module; the data-level fusion module generates full-spectrum fusion data; the feature-level fusion module concatenates core features into a comprehensive feature set; the decision-level fusion module inputs the comprehensive feature set into a hybrid model; and the trained fusion detection model processes the comprehensive feature set to output the authenticity identification result of the wheat flour sample to be tested. This invention, through a three-level spectral fusion architecture and a dynamic adaptive adversarial mechanism, can effectively achieve high-precision and high-robust identification of wheat flour authenticity.

The invention discloses a DR focus segmentation method and system based on a dynamic loss optimization function, and the method comprises the steps: collecting image data in an IDRiD data set, and carrying out the preprocessing of the image data, and obtaining the preprocessing data; based on the preprocessed data, a visual SAM large model is constructed, a loss function weight optimization strategy is designed, and an RTSAM segmentation model is obtained; segmenting the DR eye fundus image based on an RTSAM segmentation model to obtain a DR focus segmentation result; and based on the DR focus segmentation result, providing an automatic identification and segmentation visualization result of the DR focus. Aiming at the problems of non-uniform lesion distribution, non-uniform lesion occurrence frequency and the like existing in a DR color fundus image, the segmentation effect is poor, so that a dynamic loss optimization strategy is used, weight distribution in a loss function is adaptively adjusted according to a model segmentation index, and the attention of a model to rare lesion and small lesion areas is enhanced.

The invention relates to the technical field of intelligent operation and maintenance of power equipment, and discloses an AmRMR-based transformer adaptive fault diagnosis method and system, and the method comprises the steps: obtaining the detection data of dissolved gas in transformer oil, and constructing a candidate ratio feature set; respectively carrying out standardization processing on the key gas concentration characteristics and the candidate ratio characteristic set; performing discretization operation on the standardized candidate ratio feature set; performing redundancy compression and information contribution evaluation on the discretized candidate ratio features based on an AmRMR algorithm, and outputting a key ratio feature set; and inputting the key gas concentration characteristics and the key ratio characteristic set into a pre-constructed DSD-DQN fault diagnosis model, and outputting a transformer fault diagnosis result. According to the invention, the highest engineering risk that the fault is misjudged to be normal can be effectively avoided, the recognition capability of minority samples such as serious faults is remarkably improved, and transformer fault diagnosis considering engineering safety and diagnosis accuracy is realized.

The invention discloses a small-diameter tube welding defect detection method based on space guide feature selection. A used model comprises a backbone network, a pixel decoder and a query decoder. A multi-scale feature map is extracted from an input image through a backbone network, the multi-scale feature map is fused from top to bottom and from bottom to top in a pixel decoder, and a large-scale pixel decoding feature map, a secondary large-scale pixel decoding feature map, a secondary small-scale pixel decoding feature map and a small-scale pixel decoding feature map are obtained; the object query vector interacts with the small-scale pixel decoding feature map, the secondary small-scale pixel decoding feature map and the secondary large-scale pixel decoding feature map in sequence in a query decoder to obtain a prototype query vector; and interacting the prototype query vector with the foreground-enhanced large-scale pixel decoding feature map to obtain a welding defect detection result. Aiming at the characteristics of non-uniform gray distribution, smooth edge and the like of an X-ray small-diameter tube welding image, the backbone network extracts features from multiple angles, and the detection precision is improved.

The application provides a kind of intelligent construction site safety monitoring and abnormal behavior detection method, it is related to behavior identification technical field, including: based on video frame data, through target detection algorithm, personnel and cigarette in the construction site multi-modal data are positioned, to output first detection result;Based on the first detection result, video frame data and construction site multi-modal data in the preset time period, utilize 3D CNN to combine attention mechanism, to output second detection result;Based on the second detection result and construction area security level information, by isolated forest and self-encoder algorithm, get comprehensive abnormal score.

This invention discloses an automatic classification model and method for diabetic retinopathy. The automatic classification model includes convolutional input, feature extraction, and classification output. The convolutional input consists of 7×7 convolutions and max pooling, responsible for resizing the input image to extract shallow features. Feature extraction includes four alternately connected dense blocks and three transition layers, mainly responsible for extracting feature information of retinal lesions. The classification output consists of global average pooling and fully connected layers, responsible for outputting the classification result of diabetic retinopathy. This invention designs an attention mechanism (SC) based on retinal lesion features to increase the network's focus on lesion features and fully extract retinal lesion features. Using the lightweight network DenseNet121 as the backbone, the dense layers are improved by connecting the SC module after 3×3 convolutions, increasing the classifier's ability to express high-level semantic features and improving classification accuracy. This invention can achieve automatic classification of diabetic retinopathy levels.

This application relates to the technical field of artificial intelligence natural language processing, and discloses a data injection and security control method and system for psychological counseling dialogue. The method includes real-time acquisition of psychological assessment data of target users, extraction of key information and writing it into the user assessment file with a unified field structure; when a dialogue initiation signal is received, reading the most recently updated reference information from the corresponding user assessment file based on the user identifier and inputting it as a hidden prompt word into the counseling dialogue program; real-time reception of user text and inputting it into a preset emotion assessment algorithm to output emotion tendency and emotion intensity, and marking and caching abnormal emotion events and corresponding associated text; statistical analysis of the occurrence frequency, emotion intensity and recent assessment level of abnormal emotion events, and when preset intervention conditions are met, generating a corresponding intervention signal and sending it to the management terminal; this application has the effect of improving the response and security efficiency of the mental health assessment platform.

This invention discloses an intelligent scheduling system for water conservancy projects that integrates knowledge graphs, belonging to the fields of water conservancy engineering and artificial intelligence technology. It includes: a data access module, a graph construction module, a graph feature module, a decision generation module, and a constraint verification module. The data access module acquires multi-source heterogeneous monitoring data from the water conservancy project monitoring network; the graph construction module maps the data to a water conservancy project knowledge graph; the graph feature module performs feature propagation and aggregation on entity nodes based on a graph neural network to generate node embedding feature vectors; the decision generation module calculates attention weights based on an attention mechanism-based scheduling decision model to generate a preliminary scheduling scheme; the constraint verification module calls the engineering constraint rules in the knowledge graph to perform conflict detection. If a conflict is found, the associated entity nodes are backtracked, the attention weights are corrected, and the scheme is regenerated until a target scheduling scheme satisfying all constraints is obtained. This invention can effectively improve the compliance and global optimization capability of water conservancy project scheduling schemes.

The application discloses an industrial casting chill identification method based on an improved YOLOv11 and a double-flow network, comprising the following steps: S1, acquiring a multi-source data graph of an industrial casting chill; S2, performing target category and boundary box coordinate labeling on the multi-source data graph to acquire a labeled data graph; preprocessing the labeled data graph to acquire a preprocessed image; S3, dividing the preprocessed image according to a preset proportion to acquire a training set and a test set; S4, constructing an industrial casting chill identification model based on the improved YOLOv11 and the double-flow network; S5, constructing a total loss function of self-supervised learning, performing model training and evaluation on the constructed industrial casting chill identification model according to the training set and the test set, and acquiring an optimal industrial casting chill identification model, so as to realize detection and identification of the industrial casting chill according to the optimal industrial casting chill identification model. The application solves the problem of low industrial casting chill identification precision caused by insufficient detection performance in a low-light environment, occlusion and overlap, and dependence on a large amount of labeled data in the prior art.

The invention discloses a self-adaptive identification method and system for an abnormal state of an electric energy meter and a medium. The method comprises the following steps: acquiring basic data of the electric energy meter; inputting the basic data into an XGBoost and multi-attention fusion model for training, generating a reality model, and generating a virtual model consistent with the time step of the reality model; calculating a relative deviation rate and a trend difference degree of the electrical characteristic data in the real model and the virtual electrical characteristic data in the virtual model at the same time step, and taking the relative deviation rate as a characteristic difference degree; constructing a three-dimensional evaluation index system, and marking time steps exceeding a threshold value as difference indexes; the difference indexes are input into a bidirectional LSTM model, and the abnormal type of the electric energy meter is judged in combination with a dynamic decision boundary optimized by a Q-learning algorithm; and according to an abnormal type determination result, triggering an XGBoost and multi-attention fusion model to perform parameter updating. According to the invention, the misjudgment rate and the missed judgment probability in a complex scene are reduced.

The invention relates to the technical field of computer vision and agricultural automation, and particularly discloses a method for grading the quality of ziziphus mauritiana based on an improved YOLO model so as to solve the problems that a traditional grading method is low in efficiency, poor in precision and difficult to deal with complex backgrounds and small damage. According to the method, a BCW-YOLO model is provided by constructing a multi-angle Maoyesian jujube image data set and performing data enhancement, a bidirectional feature pyramid network (BiFPN) is integrated on the basis of YOLOv8 to improve the feature fusion capability, a context conversion attention mechanism (COT) is introduced to enhance detail capture, and a WIoUv3 loss function is adopted to optimize the positioning performance. According to the scheme, the model is remarkably superior to a traditional model in the aspects of accuracy, recall rate, mAP and other indexes, the model is particularly excellent in performance in complex background and small damage detection, an effective scheme is provided for achieving automatic and accurate grading of the ziziphus mauritiana, and the method has wide application prospects.

The present application relates to the field of computer vision, especially to a two-dimensional human arm posture estimation method based on feature fusion and attention mechanism, comprising the following steps: step one, collecting human action images, and performing image cropping through human boundary range; step two, performing human arm joint feature extraction on the cropped images through a basic network module; step three, performing feature fusion and focusing through a feature fusion and attention combination module, so as to reduce feature loss and improve effective information concentration, and obtaining an arm joint heat map; step four, obtaining position coordinates from the obtained arm joint heat map. The present application can obtain accurate two-dimensional arm joint position coordinates from human action RGB images.

The application discloses an electrochemical noise corrosion state identification and early warning method and system, and belongs to the technical field of corrosion monitoring and early warning. The method collects the potential noise signal and the current noise signal of the target component through an electrochemical workstation and calculates the noise resistance, adopts db4 wavelet for multi-layer wavelet decomposition and threshold denoising processing, extracts the kurtosis feature, the skewness feature and the wavelet coefficient feature, and constructs a corrosion state identification and early warning data set after feature screening; a fusion model including an LSTM time sequence feature extraction layer and an Attention attention weight distribution layer is constructed for training, and a corrosion state identification model is obtained; the real-time collected and processed feature parameters are input into the model for identification, the proportion of a specific corrosion state in a preset time period is counted, and a blue, orange or red graded early warning is triggered. The application realizes high-precision identification and accurate graded early warning of the corrosion state, the identification accuracy is above 85%, and reliable technical support is provided for industrial component corrosion protection.

The invention discloses a multi-modal large model-based remote sensing image-text cross-modal retrieval method and system, and the method comprises the steps: constructing a retrieval model and carrying out training, and the framework of the model comprises initial modal feature extraction, fine-grained image-text representation enhancement, group perception image-text similarity scoring function design and a multi-view contrast learning strategy; respectively extracting initial representations of the text and the image; respectively enhancing the two types of fine granularity representations; for the image, three learning tasks of feature reconstruction, scene classification and background information alignment are taken as guidance, the background information of each image block is separated from target features, and meanwhile, the relationship between targets is reinforced through image learning; key word information is emphasized for the text; and calculating correlation scores and generating the similarity of the image-text pairs. The system comprises a model building unit, a model training unit and a retrieval unit. By using the method and the device, the cross-modal retrieval precision can be improved. The method can be applied to the field of cross-modal retrieval.

This application relates to a live streaming interaction method, apparatus, live streaming system, electronic device, and computer-readable storage medium in a live streaming room. The method includes: in response to a viewer's operation of entering a recommended channel to watch a live stream, loading and displaying a live streaming aggregation interface of the recommended channel from a live streaming server; obtaining multiple recommended anchors issued by the live streaming server, and displaying a switchable live streaming room of a main anchor and the live streaming room covers of each recommended anchor on the live streaming aggregation interface; receiving the live streaming content of the main anchor's live streaming room forwarded by the live streaming server and displaying it on the live streaming aggregation interface; in response to a viewer's operation of interacting with the main anchor on the live streaming aggregation interface, executing the live streaming interaction between the main anchor and the viewer, and synchronizing the interactive content generated by the live streaming interaction on the recommended channel to the main anchor's live streaming room. This technical solution increases the number of followers of recommended anchors and increases the daily active users of the live streaming platform.

Embodiments of the present application disclose a core image analysis model training method and device, electronic equipment and a storage medium. The method comprises: performing block processing on the obtained core image, and using a mask image modeling algorithm to randomly mask the block image to obtain a masked block image; performing linear transformation on each block image and adding position information; using a self-attention mechanism to extract first image features in each unmasked block image, and using a register to predict second image features of the masked block image; using a self-supervised learning algorithm to train the model according to the first image features and the second image features, and using a knowledge distillation algorithm to optimize the model to obtain a target core image analysis model. The technical solution of the embodiments of the present application can better capture the global information and detailed structure in the core image, thereby effectively improving the accuracy and stability of feature extraction.

The application relates to the field of intelligent decision-making, and discloses a face living body detection method, device and equipment in a motion state scene and a medium. The method comprises the following steps: collecting a face image in a motion state scene, performing face detection on the face image to obtain a detected face, performing format standardization on the detected face to obtain a standardized face; performing texture coding on the standardized face to obtain coded texture of the standardized face, extracting texture information, calculating pixel change information, and constructing three-dimensional structure information of the standardized face; respectively performing feature extraction on the texture information, the pixel change information and the three-dimensional structure information to obtain texture features, pixel change features and three-dimensional structure features; performing feature fusion on the texture features, the pixel change features and the three-dimensional structure features to obtain fused features; and calculating a living body detection score of the standardized face to determine a face living body detection result of the face image. The application can improve the comprehensiveness of face living body detection in a motion state scene.

The invention provides a power distribution network development form probability prediction model construction method and system, and the method comprises the steps: collecting multi-source heterogeneous data of micro-grid nodes in a target region, and carrying out the development form type marking of the multi-source heterogeneous data, and constructing a time-space correlation data set; based on attribute correlation and time proximity of samples in the space-time association data set, dynamically compensating and repairing the missing values to generate space-time alignment samples, and training a development form probability prediction model by taking the space-time alignment samples as input and taking development form category probability distribution corresponding to the space-time alignment samples as output to obtain a development form probability prediction model; and obtaining a trained development form probability prediction model. Through the space-time decoupling layer, the feature fusion layer, the dynamic graph convolution layer and the classification output layer of the development form probability prediction model, the problem that the development form deduction prediction precision is insufficient because multi-source heterogeneous data cannot be effectively fused, space-time features cannot be distinguished, key mutation events cannot be captured and dealing with sample imbalance in the prior art is effectively solved.

The invention discloses an industrial product surface nondestructive testing method and system based on improved YOLOv8, and relates to the technical field of computer vision and image processing. In a YOLOv8 backbone network, standard 3 * 3 convolution is replaced by a space channel selective kernel convolution module, and the nondestructive testing of the surface of an industrial product is realized by performing space and channel reconstruction on features and dynamically adjusting a receptive field. Redundant features are suppressed, and the modeling capability for complex textures is enhanced; a coordinate-channel-space joint attention module is embedded in front of a detection head classification branch, and extraction and focusing of small defect features are enhanced through coordinate, channel and space attention mechanisms; structural upgrading is carried out on the feature fusion part of the network, the detection scale is expanded to a P2 layer, and the capacity of capturing tiny targets is improved. In addition, multi-scale feature fusion and non-maximum suppression and weighted frame fusion in a post-processing stage are combined, so that the positioning precision and the result stability of the detection frame are further improved.

The application relates to the technical field of voltage compensation control, in particular to a power distribution line dynamic compensation adjustment method for preventing voltage fluctuation. The method comprises the following steps: acquiring voltage deviation degrees and reactive power variation synchronization degrees between each trunk node and each compensation point at each moment; acquiring limit coefficients of each trunk node, and dividing an alternating current power grid area into reactive power compensation coverage sub-areas dominated by each compensation point; calculating line losses of trunk lines in each reactive power compensation coverage sub-area, and acquiring multi-time scale estimation information; acquiring compensation parameters of reactive power compensation devices in each reactive power compensation coverage sub-area, and then correcting compensation capacities of the reactive power compensation devices to dynamically compensate the power distribution line. The application aims to reduce active power loss and voltage drop and improve overall power quality.

The invention provides a data center park power transmission line inspection system and method, and relates to the technical field of computer vision, and the method comprises the steps: carrying out the multiple continuous deformable convolution of a remotely collected power transmission line image, carrying out the preliminary feature extraction, and carrying out the deep feature mining through spatial pyramid pooling and triple attention, multiple layers of feature maps with different scales are obtained through co-extraction; performing multi-stage feature fusion on the feature maps of all scales to output fusion information of four scales; and a plurality of detection heads are arranged to detect the different fusion information to obtain the target category of the fusion information and the position of the fusion information in the power transmission line image. The method has the beneficial effects that attention on fine-grained deep information and transmission of effective information are enhanced through deep feature mining, and the robustness and generalization of the model are improved; through multi-stage feature fusion and detection, interaction between deep semantic information and shallow position information is enhanced, and the detection precision and accuracy of the system are improved.

The invention relates to the technical field of campus security and protection management, in particular to a smart campus intelligent security and protection early warning system, which comprises a data acquisition module used for acquiring campus security and protection data through a multi-dimensional security and protection device and a patrol robot; the vulnerability thermodynamic diagram construction module is used for constructing a risk simulation model to calculate the security vulnerability coefficient of each geographic grid unit of the campus, dividing risk levels and generating a vulnerability thermodynamic diagram; the early warning module is used for carrying out multi-dimensional analysis, identifying a security anomaly type and triggering corresponding early warning; the patrol scheduling module is used for dynamically planning patrol routes of the patrol robots, and when early warning is triggered, the patrol robot closest to an early warning area is scheduled to go to the site for checking; and the role collaboration module is used for pushing differentiated tasks to different security roles based on the risk level distribution and the early warning type of the vulnerability thermodynamic diagram. Therefore, the problems of single data dimension, risk perception lagging, resource scheduling stiffness, low efficiency of multi-role collaboration and the like are solved.

The application relates to the technical field of time sequence knowledge graph, in particular to a representation learning method of a time sequence knowledge graph based on local-global feature fusion. The method comprises the following steps: dividing an entire time axis into multiple time points, and constructing a time sequence knowledge graph based on the time points; extracting local space-time features based on the time sequence knowledge graph by using a local encoder, and extracting global features by using a global encoder; fusing the local space-time features and the global features by using a gate mechanism adaptive weighting method; predicting entities by using a ConvTransE model based on the fused features by using a decoder, and predicting relationships by using a ConvTransR model based on the fused features by using the decoder; and feeding back the prediction results to a loss function driving end-to-end training. By combining the space-time features and global semantic information in the knowledge graph extracted by the local encoder and the global encoder, the application can more accurately capture the dynamic changes of entities and their relationships in the time dimension.

The application discloses a kind of embodied large model construction methods for adaptive grasping task, it is related to embodied intelligent technical field.The method includes obtaining training dataset;Embodied large model for adaptive grasping task is constructed, including constructing visual feature extraction module, constructing language feature extraction module, constructing unified semantic representation module, constructing cross-modal transfer module, constructing dynamic modal scheduling module, constructing action output module;Training dataset is used to train the embodied large model, and the embodied large model for adaptive grasping task that training is completed is obtained.The embodied large model constructed in practical application can improve the accuracy of grasping action generation, environmental adaptability and job stability of automated grasping equipment in complex grasping scene, effectively reduce the dependence degree of artificial demonstration and repeated parameter adjustment.

The application relates to the field of small sample target detection, in particular to a two-stage small sample target detection method based on an optimized CBAM attention mechanism, and comprises the following steps: training a two-stage target detection network Faster-RCNN by using a base class data set to obtain a base class detection model; freezing parameters of a feature extraction backbone network in the base class detection model; optimizing a CBAM attention mechanism module; placing the optimized CBAM attention module in the feature extraction backbone network to construct a detection network, then inputting a new class small sample data set with a small amount of labeled information to fine-tune parameters of a detection head part of the detection network; and inputting a to-be-detected data set into the detection network to obtain a detection result. Compared with the prior art, the application has the advantages of inhibiting the influence of unimportant spatial information, improving the attention degree of important spatial information, enhancing the sensitivity to different scale features, and having strong generalization ability and robustness and the like.