138 results about "Linear complexity" patented technology

A hybrid approach involving the mix of communication patterns and passive fingerprinting is used to identify unknown device types, manufacturers, and models of devices of digital control systems. ANDVI implementation maps the identified devices to their known vulnerabilities. According to one example, to identify how interdependence among existing atomic vulnerabilities may be exploited by an adversary to stitch together an attack that can compromise the system, model-checking based A2G2V is employed. According to another example, an A2G2V algorithm uses existing model-checking tools, an architecture description tool, and code to generate an attack-graph that enumerates the set of all possible sequences in which atomic-level vulnerabilities can be exploited to compromise system security. In yet another example, identification of label-cuts within an attack-graph automatically identifies a set of critical-attacks that, when blocked, renders the system secure. A linear complexity approximation utilizing SCCs helps identify the minimum label-cut representing a critical-attacks set.

The invention discloses a medium and long term weather prediction system based on a physical law correction deep learning technology, and the system comprises a high-resolution data input module, a multi-resolution network construction module, a sub-graph processing module based on a graph neural network, and a linear attention processing module. The system comprises an advanced attention processing module used for reducing secondary calculation complexity to linear complexity, a feature fusion and adaptive fusion module, a multi-level decoding and physical corresponding area optimization correction model, a rolling forecasting module and a multi-step training fine tuning module based on DORA. The medium and long term weather prediction system based on the physical law correction deep learning technology can process multi-resolution input data, and realizes high-precision and long-time weather prediction. Furthermore, global weather forecast with the resolution ratio of 16 days and 13 kilometers can be achieved, and meanwhile refined forecast with the higher resolution ratio is achieved in China.

The invention discloses a physical information neural network-driven ocean island typhoon scene generation method and system. The method comprises the steps of collecting multi-source heterogeneous meteorological data and performing space-time alignment preprocessing; constructing a coarse-scale space-time probability prediction model, capturing space correlation of meteorological elements by using a graph topology learning network, efficiently processing long-time-sequence dependence of typhoon evolution by integrating a state space model with linear complexity, and generating a probabilistic typhoon scene with coarse resolution through a multivariable joint distribution probability model; further constructing a physical downscaling model, taking a coarse-scale prediction result as condition input, and performing physical consistency downscaling on a coarse-scale scene by embedding an atmospheric fluid mechanics equation in a loss function as a physical hard constraint; and finally, outputting a high-resolution typhoon scene with probability reliability and physical authenticity.

The invention discloses a medical image fuzzy boundary segmentation method based on an edge perception Mama network. The method aims at solving the problems that a camouflage pathological structure in a medical image is visually similar to surrounding healthy tissues, so that segmentation is difficult, and clinical deployment is difficult due to secondary calculation complexity of an attention mechanism in a traditional camouflage target detection (COD) method. According to the invention, a boundary guiding module inspired by COD is creatively combined with linear complexity state space modeling, and an E-Mama edge guiding framework is provided. The framework adopts an encoder-edge guide-decoder structure, and accurate boundary description is realized while the computational efficiency is kept. The system comprises five core components: a Mama encoder; an adaptive fusion processing module (AFM); an edge detection module (EDM); an edge guidance module (EGM); the invention relates to a Mama decoder. Experiments show that advanced segmentation precision is realized on a plurality of medical data sets, and meanwhile, the calculation complexity is reduced from O (n2) to O (n), so that clinical deployment becomes possible.

The invention discloses an optic disk atrophy arc image segmentation method based on full supervision, and particularly relates to the technical field of medical equipment, and the method comprises the following steps: S1, multi-scale feature extraction, S2, local-global feature enhancement, S3, adaptive feature fusion, S4, gradient flow optimization, S5, mixing of a loss function, and S6, data enhancement and generalization optimization. According to the method, the long-range dependence of the optic disk and the atrophic arc is globally modeled through the PVTv2 encoder, the detail segmentation precision of the optic disk atrophic arc at the blood vessel crossing and fuzzy boundary is remarkably improved in combination with the multi-scale attention and dilated convolution of the feature enhancement module, cross-level feature fusion is realized with linear complexity based on the Mama decoder, and the accuracy of feature fusion is improved. Efficiency and structure coherence are considered; the weight is balanced by the mixed loss function, and early lesion missing detection is reduced; a data enhancement and residual module enhances the generalization ability of the model, and a high-precision and efficient quantitative analysis tool is integrally provided for early screening of blind eye diseases such as pathological myopia and glaucoma.

The invention discloses an underwater image semantic segmentation method and device based on a lightweight double-flow Mama network, and a storage medium, relates to the technical field of underwater image processing, and solves the problems of poor environmental adaptability, low modal fusion efficiency, heavy model and high calculation overhead in the existing underwater image semantic segmentation technology. According to the method, a double-branch encoder is constructed, the double-branch encoder composed of an image encoder and a text encoder is adopted, and a visual feature map and a semantic feature vector in a preprocessed underwater image and text description information are extracted respectively; a cross-modal Mama module is adopted to carry out deep fusion on a flattened image feature sequence and text features in the module, the cross-modal Mama module adopts a Mama block with linear complexity, and continuous guidance and progressive enhancement of text semantics are realized in combination with a multi-level gating fusion mechanism and residual connection. And the recognition capability of underwater fuzzy and shielded targets is remarkably improved, and meanwhile, the calculation efficiency is remarkably improved.

The invention provides an anti-interference serial port communication method and system for an industrial-grade wireless module, and the method comprises the steps: collecting interference data in the serial port communication of the industrial-grade wireless module in real time based on a multi-dimensional interference feature collection module, and forming a multi-scale interference feature vector; inputting the multi-scale interference feature vector into a Mamba communication state space model fused with a CRC (Cyclic Redundancy Check) mechanism, and predicting a potential interference mode through linear complexity time sequence modeling; based on an adaptive weight transmission optimization function, carrying out dynamic weighting adjustment on the transmission parameters in the potential interference mode; wherein the transmission parameters comprise a baud rate, a data bit length, a verification mode and a retransmission strategy. According to the invention, hardware cost does not need to be increased, and the cost of anti-interference serial port communication is reduced; meanwhile, based on an adaptive weight transmission optimization function, dynamic weighting adjustment is carried out on transmission parameters in a potential interference mode, and interference dynamic changes are tracked in real time to carry out anti-interference communication.

The invention relates to the technical field of protein prediction, and discloses an efficient protein stability prediction method for selective state space modeling. According to the method, the BiMama core module and the CoGNN core module are adopted, and limitation of an existing method on calculation efficiency and multi-scale information processing is broken through in a mode of combining selective state space modeling and the collaborative graph neural network. According to the invention, a local sampling strategy based on a k-hop sub-graph is provided, and efficient calculation is realized by focusing a key environment around a mutation site; a bidirectional selective state space model is adopted to model a long-range dependency relationship with linear complexity, and the calculation bottleneck of a traditional Transform architecture is effectively overcome. The gating fusion module developed by the invention can adaptively integrate long and short range features, so that the model can flexibly adjust a feature combination strategy for different types of mutations, and the accuracy and efficiency of protein delta delta G prediction are improved.

The invention relates to the technical field of sleep stage classification, and particularly provides a sleep quality monitoring method fusing multi-scale features. The method comprises the following steps: preprocessing input signals, and inputting the input signals into a multi-mode sleep signal analysis network after ensuring that the signal amplitudes are consistent; capturing features of a plurality of time scales through a multi-granularity feature learning module, and carrying out feature fusion; integrating spatial information by using a spatial-temporal feature enhancement module, and mapping the three-dimensional features into a two-dimensional time sequence feature sequence; on the basis of a time context module of Mama, long-distance time dependence is modeled with linear complexity, features are averagely aggregated along the time dimension through a classification module, a classification result is output through a full connection layer and an activation function Softmax, and the method improves the accuracy of sleep stage classification while ensuring the high efficiency of sleep stage classification.

The invention relates to the technical field of intelligent agriculture and computer vision, in particular to a small-sample high-density chicken counting framework based on a deep learning Mamba structure, which comprises a feature extraction network, a support-query enhancement module and a decoder. According to the method, a multi-scale feature extraction network based on a residual block (ResNet Block) is introduced, so that local detail information is effectively reserved; and then, a support-query enhancement module is constructed by introducing a Mamba structure, and context interaction between support features and query features is effectively enhanced by utilizing the long sequence modeling capability of linear complexity of the support-query enhancement module, so that the problems of individual overlapping and boundary fuzziness in a high-density chicken flock scene are solved. Experimental results on a PoultryCount real breeding data set show that the mean absolute error (MAE) and the root-mean-square error (RMSE) of the method are reduced compared with those of an existing method, and the chicken counting precision and generalization ability under the condition of a small number of labeled samples are remarkably improved.

The invention relates to the technical field of quantum computing, and provides a linear complexity quantum state preparation method based on tensor decomposition and a quantum circuit construction system.The high-dimensional tensor is decomposed into a series of low-rank core tensors through continuous singular value decomposition, each core tensor in a core tensor sequence is expanded into a unitary matrix, and the unitary matrix is used as a quantum circuit; the unitary matrix sequence is mapped to quantum lines coupled using adjacent qubits, and the quantum lines are run to prepare a target quantum state. Based on this, the line generated by the method can approximately or accurately prepare a target quantum state only by coupling adjacent quantum bits, and is perfectly adaptive to quantum chips of linear or grid topologies such as superconducting and semiconductor quantum dots and the like.

A single-target tracking method and device based on a state space model and attention and a medium are characterized in that firstly, hierarchical feature extraction is carried out, hierarchical modeling is carried out on a template and a search area through a mixed attention mechanism and the state space model, and a time sequence token is introduced into each frame at the final stage of hierarchical feature extraction to carry out target information aggregation; and aggregating the time sequence tokens by using a state space model to realize cross-frame information propagation, multiplying the obtained time sequence tokens by corresponding frame feature maps to serve as feature enhancement, and finally predicting a bounding box of a target object by a prediction head to realize target representation. According to the method, the state space model and the attention mechanism are combined for single-target tracking, especially target tracking under a large-resolution video, the state space model is introduced into the field of single-target tracking, the advantage of linear complexity of the state space model is brought into full play, meanwhile, the disadvantage of limited retrieval capacity of the state space model is made up, and high-resolution video tracking is achieved. And finally, the consumption of computing resources is reduced while accurate tracking is realized.

The invention relates to the technical field of medical data processing, in particular to a data analysis and processing system based on clinical hearing detection, and aims to solve the problems that dynamic characteristics of auditory nerve response cannot be systematically described, high-dimensional multi-modal feature vectors cannot be constructed, and the data analysis and processing efficiency is low in the prior art. A stable and reliable input basis cannot be provided for intelligent classification, anomaly detection and individualized analysis, and the sensitivity and specificity of hearing state recognition are reduced; the multi-dimensional hearing feature extraction module fuses time domain, frequency domain and non-linear complexity analysis methods, systematically depicts dynamic features of auditory nerve response, constructs a high-dimensional multi-modal feature vector by integrating three-dimensional features, remarkably improves sensitivity and specificity of hearing state recognition, and improves the accuracy of hearing state recognition. The feature representation method provides a stable and reliable input basis for subsequent intelligent classification, anomaly detection and individualized analysis.

The invention discloses a trajectory planning method based on a multi-stage optimization strategy and a mixed diffusion model. The method comprises the following steps: fusing multi-source sensor data and extracting features; screening and fusing intention anchor point tracks; and optimizing the fine-grained trajectory based on the Transform-Mamba mixed diffusion model. According to the method, an intention anchor point track screening mechanism is provided, the anchor point track matched with the scene is selected from the compact vocabulary through the dynamic screening module and fused with the static prior anchor points, the calculation complexity is reduced, the diversity of the initial track and the scene consistency are guaranteed, and the problem that a traditional fixed anchor point set is insufficient in flexibility is solved. According to the method, a mixed diffusion decoder is designed, space environment dependence is efficiently modeled through a cross attention mechanism, a bidirectional Mama module captures time sequence motion dependence with linear complexity, global perception ability and long sequence modeling efficiency are both considered, and safety and dynamics feasibility of generated tracks are ensured.

The invention belongs to the technical field of machine state prediction, and particularly relates to a fault diagnosis method and system based on a coding and decoding attention mechanism enhanced pulse neural network, and the method comprises the steps: collecting a device vibration signal, and segmenting the device vibration signal into time sequence segments; converting each fragment into a two-dimensional time-frequency image by using S transformation to construct a data set; the image is input into a lightweight hybrid network model to extract deep features, the model is composed of a pulse convolution encoder and a pulse efficient additive attention mechanism encoder which are alternately cascaded, and the pulse convolution encoder extracts local features through pulse neurons and depth separable convolution; the global context dependency is modeled by adopting an efficient additive attention mechanism with low linear complexity; and finally, fault classification is completed based on deep features. Through fusion of pulse calculation and an efficient attention mechanism, the feature extraction capability and diagnosis precision are improved while the parameter quantity of the model is remarkably reduced, and the method is particularly suitable for being deployed on edge equipment with limited resources to realize fault diagnosis.

The invention relates to the technical field of computer vision and document understanding, and discloses a table recognition method based on structure perception coding and state space decoding. The method comprises the steps of firstly extracting table image features; then, features are enhanced through structure perception coding, two-dimensional position coding and row and column embedding are superposed to form grid prior, and serialized visual representation is constructed to improve modeling of a global structure and a row and column alignment relation and reduce redundant calculation. In the decoding stage, state space decoding with linear complexity is adopted, and a table structure mark is generated through autoregression; when a cell start mark is detected, synchronous prediction of cell positioning and content decoding is triggered, and joint optimization is carried out in training. And finally, carrying out fusion and consistency verification on the structure and content results, and outputting a standardized structure and content alignment grid. The method has high recognition precision, robustness and reasoning efficiency in complex scenes of merging cells, invisible boundaries, irregular layout, super-long sequences and the like.

The invention relates to a medical image segmentation system and method based on context and frequency guidance and a storage medium, and the system employs an encoder-decoder architecture, and integrates three core modules: a visual state space module which captures global context information with linear complexity based on a selective state space model; the frequency-guided representation module is used for explicitly separating and enhancing structure and boundary characteristics through frequency domain transformation and complex weight modulation; and the multi-scale adaptive context aggregation module is used for integrating multi-scale semantics through parallel multi-branch convolution and dual Top-K sparse attention and focusing a key region. According to the method, the technical problems that global modeling and boundary details are difficult to consider, the calculation complexity is high and the adaptability to multi-scale targets is poor in the prior art are solved, the segmentation precision and the boundary description accuracy are remarkably improved in various medical image segmentation tasks such as heart MRI, polyp, skin lesion and pathological sections, and the medical image segmentation efficiency is improved. And meanwhile, the calculation efficiency is ensured.

The invention discloses a method for designing a double-iteration Turbo equalizer of an OTFS-MIMO-ISAC system, and the method comprises the steps: carrying out the sparse processing of covariance matrix inversion, and reducing the matrix inversion cost of conventional MMSE equalization through linear complexity derivation; the sparse linear system is solved through the TFQMR algorithm to reduce the complexity of the first iteration, and a sparse approximate inverse matrix is generated in the subsequent iteration through the FSPAI algorithm to further realize the balance between the calculation efficiency and the performance; and dynamically optimizing channel information by adopting an iterative channel estimation method and utilizing MMSE (Minimum Mean Square Error) equilibrium feedback. Through outer iteration and inner iteration, the symbol estimation precision is gradually improved, meanwhile, the calculation complexity is reduced, and a low-complexity and high-precision equalization and channel estimation solution is provided for a high-mobility wireless communication system.

The invention discloses a double-flow fault diagnosis method based on state space modeling and dynamic edge context awareness, the method constructs a time domain state flow and a time frequency edge flow, and the specific implementation process is as follows: the time domain state flow is subjected to state space modeling under a Transform framework to form a MambaT encoder, long and short term dependence is efficiently captured with linear complexity, and the time domain state flow is subjected to dynamic edge context awareness; time domain sensitive features are accurately represented; the time-frequency edge stream converts vibration data into a two-dimensional time-frequency image through wavelet transformation, a differentiable edge sensitive mask is designed to adaptively separate time-frequency image impact characteristics and background noise, boundary content perception adaptive filling is designed to maintain image boundary topology integrity, and the two parts cooperate to form dynamic edge context perception convolution, so that the image boundary topology integrity is improved. Time-frequency edge features are effectively extracted; and a channel enhancement and channel gating module is designed to establish a dynamic channel interaction gating mechanism, so that time domain-time frequency domain double-current cascade characterization is realized, and the fault discrimination capability is remarkably improved. Experiments prove that the method has higher fault diagnosis performance.

The embodiment of the invention provides a text map generation method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining an initial U-shaped neural network model used for image denoising processing in a text map generation process, and the initial U-shaped neural network model comprises a plurality of groups of attention structures; replacing the multiple groups of attention structures with a target model structure with linear complexity; performing model distillation on the U-shaped neural network model after the structure replacement based on the initial U-shaped neural network model to obtain a target U-shaped neural network model; and according to the target U-shaped neural network model, carrying out image denoising on the text map to obtain a target image. According to the embodiment of the invention, the target model structure with linear complexity is adopted to replace multiple groups of attention structures, the calculation time complexity of the attention module is reduced to linear complexity, and the target U-shaped neural network model is obtained, so that the text map can generate the image with the resolution greater than 2k.

The invention belongs to the technical field of medical image segmentation, and relates to a dual-path fusion neural network for prostate precise segmentation and a segmentation method, the neural network constructs a dual-path decoupling encoder architecture based on an nnU-Net framework, captures fine anatomical structure and local texture information through a context sensing residual encoder of a local path, and obtains a dual-path fusion neural network for prostate precise segmentation. A long-range dependency relationship is modeled with linear complexity through a visual state space module of a global path, a double-flow alignment gating module is designed to realize self-adaptive alignment and fusion of cross-path features, and model training is optimized in combination with a mixed loss function and a depth supervision strategy; according to the method, the limitation of an existing segmentation model in the aspect of local structure and global semantic integration is solved, and the segmentation precision, the boundary goodness of fit and the generalization performance of the focus and gland region in the prostate MRI image are improved.

The invention relates to the technical field of electric digital data processing, in particular to a target analysis method and system based on a complex network, and the method comprises the following steps: storing an original view; constructing a coarsened view; constructing a line graph conversion view; embedding and fusing the three views through a general graph converter, and inputting the three views to a classifier of a target analysis model based on a complex network; obtaining a classification result of the target analysis model based on the complex network, and calculating a cross entropy loss function; and the classifier performs back propagation, and repeatedly updates the weight and bias of the target analysis model based on the complex network for multiple rounds by using a gradient descent method until an accurate classification result of the target to be analyzed is obtained. According to the method and the system provided by the invention, linear complexity analysis is realized, and the problem of performance reduction caused by distribution offset is relieved, so that more efficient and accurate target analysis in a complex network is realized.

The invention discloses a question and answer method and system based on large model causal graph discovery and causal graph enhanced reasoning, and belongs to the technical field of natural language processing and artificial intelligence. The method comprises the following steps: firstly, carrying out hierarchical expansion from an initial root node by adopting a breadth-first search strategy, and efficiently constructing a reliable causal graph conforming to directed acyclic graph constraints with linear complexity in combination with real-time loop detection; in the question and answer stage, key entities in user questions are analyzed, and causal paths connecting the key entities are retrieved in a causal graph; and taking the retrieved structured causal path as a constraint condition to be injected into a decoding process of a large-scale language model, and generating a natural language answer which is strict in logic and can trace reasoning steps. According to the method, the resource consumption of large-scale causal discovery is remarkably reduced, the accuracy and interpretability of the answers of the questions and answers are effectively improved, and technical support is provided for medical treatment, finance and other scenes needing high-reliability reasoning.

The invention provides a novel spatial micro-motion target identification method based on a radar network system, and the method comprises the following steps: building a target micro-motion model and a signal model, and constructing a data set of an RCS sequence, a time Doppler image and an HRRP sequence for the spatial micro-motion target identification of a networking radar system; constructing an adaptive feature fusion sub-network; xCA and LA-Att modules with linear complexity are introduced, fine-grained features are effectively extracted, self-adaptive weighted fusion of multi-domain features is realized by adopting a CAFM module, and information useful for recognition is effectively enhanced; and designing a time-space modeling sub-network based on BiGRU and double-graph fusion. According to the method, multi-band and multi-view target features can be obtained, and more accurate information is provided for identification of a spatial micro-motion target; according to the method, the feature information of each radar in the heterogeneous radar network is fully utilized, the recognition precision of 96% or above can still be obtained under the condition that the signal-to-noise ratio is 0dB, and the effectiveness and robustness of the method are obvious.

The invention provides a remote sensing image target detection method using a scale-sensitive Mama network, which comprises the step of preparing a training set and a test set of a DIOR data set, and specifically comprises the following steps: carrying out SSMNet network training by using images in the training set to generate a training model; storing the training model in a local folder, testing the effect of the training model by using the images in the test set, and if the training model is satisfied, storing the training model as a satisfied training model; and storing the satisfactory training model in a local folder, and testing an unlabeled remote sensing image by using the satisfactory training model. The invention provides a scale sensitive Mama network which comprises an area intersection-to-parallel ratio (AIoU) loss function, a remote sensing Mama block (RSMB) and a sparse feature fusion module (SFFM). The AIoU improves the scale sensitivity and the small target convergence speed; the RSMB models a global context with linear complexity; and SFFM is fused with deep and shallow features to enhance small target detection. The model significantly improves the remote sensing image target detection effect.

The invention relates to the field of image segmentation, in particular to an image segmentation method and device based on an SM-UNet image segmentation model. The invention provides a novel encoder and decoder Mama framework, a sliding window is applied to a UNet segmentation model based on Mama to establish an SM-UNet image segmentation model, Swin Mama UNet (SMUNet) is used for automatic medical image segmentation, and the advantage of Mama linear complexity and the cross-window interaction capability brought by the sliding window are mainly combined to effectively optimize the structure of a standard U-shaped framework.

An efficient and scalable attention model is provided, which can be referred to as multi-axis attention. An exemplary implementation can include two aspects: blocked local attention and spread global attention. These design choices enable global-local spatial interaction for any input resolution with only linear complexity. This disclosure also presents a new architectural element by effectively fusing convolution with the proposed multi-axis attention model. In addition, this disclosure proposes a simple hierarchical vision backbone by simply repeating basic building blocks across multiple stages, an exemplary implementation of which can be referred to as MaxViT. In particular, MaxViT is able to "see" globally throughout the entire network, even in earlier, high-resolution stages.

The invention provides an efficient single-stage super-resolution method and a remote sensing image super-resolution system for RGB remote sensing images, and the method integrates Fourier position coding guided geometric node modeling, hypergraph high-order relation modeling and a Mama long-range dependence modeling mechanism based on a state space model. And consistent modeling of a multi-scale structure is realized while the linear calculation complexity is kept. In the method, FOPE is introduced as explicit geometric priori to guide adaptive screening of key nodes on global, regional and local scales; then, an FOPE enhanced hypergraph network is constructed, and high-order space association among multiple nodes is modeled through hyperedge; according to the method, a two-stage Mama aggregation module is further designed, and intra-node feature consistency modeling and inter-node long-range spatial dependency modeling are completed with linear complexity; and finally, a high-resolution RGB remote sensing image is generated through a lightweight direct up-sampling module.

The invention relates to an RGB-T target tracking method and system based on spatio-temporal state evolution, and the method comprises the steps: constructing a double-branch RGB-T target tracking model, enabling RGB and TIR branches to share the weight of a ViT encoder, employing an iterative processing frame, and transmitting spatio-temporal context information between frames through an updatable context memory Tokens. In each iteration, after RGB and TIR modal features are extracted respectively, cross-modal time sequence context modeling is carried out through a modal perception time sequence Mamba module, and the module realizes long-time target perception representation learning through a cross-modal coupling state transition mechanism and a prompt guide strategy; and carrying out intra-modal and inter-modal feature fusion in a spatial dimension through a cross-modal Mama aggregation module, and finally outputting a target position through a prediction head. According to the method, lasting cross-modal state evolution can be realized with linear complexity, the spatial-temporal characteristics of visible light and infrared light are effectively fused, and the tracking robustness and accuracy are improved in complex scenes such as rapid target movement, shielding or modal degradation.

The present application relates to the field of artificial intelligence. Provided in the present application are a method and system for implementing a self-attention mechanism having a linear complexity, and a device and a medium. The method comprises: performing positional encoding on sequence data to be processed, and performing feature mapping to apply the positional encoding to a query matrix and a key matrix; on the basis of the query matrix and the key matrix, performing calculation to obtain a low-rank query matrix, a low-rank key matrix and a low-rank generalized inverse matrix; augmenting a value matrix to obtain an augmented matrix, and on the basis of the augmented matrix, the low-rank key matrix, the low-rank generalized inverse matrix and the low-rank query matrix, obtaining a sequence self-attention matrix; and transposing the sequence self-attention matrix, using the transposed matrix as sequence data to repeat the foregoing steps, using the newly obtained sequence self-attention matrix as a feature self-attention matrix, and on the basis of the sequence self-attention matrix and the feature self-attention matrix, obtaining a self-attention matrix having a linear complexity. The present application implements linear self-attention, better maintains the model performance and has relatively good expansibility.