Neuromorphic Visual Computing Method and System for Fast Recognition of Dynamic Scenes

Abstract

The present invention relates to the field of computer vision technology, and discloses a neuromorphic vision computing method and system for rapid recognition of dynamic scenes, including: obtaining a quantum dynamic image of a dynamic scene, and identifying quantum polarization characteristics to separate interference features of the quantum dynamic image to obtain a clean information sequence; performing optomagnetic conversion on the clean information sequence to obtain a pulse feature map, identifying the quantum correlation features of the pulse feature map, and performing quantum feature fusion on the pulse feature map to obtain a fusion feature tensor; performing topological mapping on the fusion feature tensor to obtain a mapped feature, calculating the scene attention weight of the dynamic scene, separating regions of the quantum dynamic image to obtain a target analysis region; collecting the quantum spatio-temporal pulse sequence of the target analysis region, performing multi-level decomposition to obtain a decomposed pulse sequence, constructing a comprehensive correlation map of the dynamic scene, and performing target recognition on the dynamic scene. The present invention can achieve flexible recognition of high-complexity dynamic scenes.

Description

Technical Field

[0001] The present invention relates to a neuromorphic vision computing method and system for rapid recognition of dynamic scenes, belonging to the technical field of computer vision. Background Art

[0002] In today's intelligent vision application field, the rapid recognition of dynamic scenes is crucial. From the real-time capture of abnormal behaviors in intelligent security monitoring to the timely response of autonomous driving systems to the ever-changing situations on the road, the level of dynamic scene rapid recognition technology directly affects the reliability and security of these applications, and thus is related to the social operation efficiency and people's quality of life. Accurately and quickly recognizing dynamic scenes can timely respond to various emergencies, avoid potential dangers, and improve the overall performance of the system.

[0003] Currently, for the rapid recognition of dynamic scenes, traditional vision computing technologies are mostly adopted. This technology usually uses image acquisition devices such as cameras to obtain a sequence of scene images, and uses image processing algorithms to extract and analyze the features of the images. Then, based on the analyzed feature information, a model (such as a convolutional model) is constructed, and some fixed recognition rules and processes are set in combination with past experience to perform the recognition task. However, dynamic scenes often have a high degree of complexity and real-time dynamic variability. For example, in actual application scenarios, there are complex situations such as diverse motion patterns of target objects, complex lighting changes, and background dynamic interference. In the face of such scenes, due to the fixed recognition mode of traditional vision computing technologies, the recognition system cannot adaptively adjust, resulting in poor recognition effects of this method for complex dynamic scenes. Summary of the Invention

[0004] The present invention provides a neuromorphic vision computing method and system for rapid recognition of dynamic scenes, and its main purpose is to flexibly recognize highly complex dynamic scenes.

[0005] To achieve the above object, the neuromorphic vision computing method for rapid recognition of dynamic scenes provided by the present invention includes:

[0006] Deploy a visual camera array in the dynamic scene to be rapidly recognized to obtain the quantum dynamic image of the dynamic scene, identify the quantum polarization characteristics of the quantum dynamic image, and based on the quantum polarization characteristics, separate the interference features of the quantum dynamic image to obtain a clean information sequence;

[0007] Construct an optical signal coding matrix of the clean information sequence, use the optical signal coding matrix to perform optical-magnetic conversion on the clean information sequence to obtain a pulse feature map, identify the quantum correlation characteristics of the pulse feature map, and based on the quantum correlation characteristics, perform quantum feature fusion on the pulse feature map to obtain a fusion feature tensor;

[0008] Perform a topological mapping on the fused feature tensor to obtain mapped features. Using the mapped features, calculate the scene attention weights of the dynamic scene. Based on the scene attention weights, perform region separation on the quantum dynamic image to obtain a target analysis region;

[0009] Collect the quantum spatio-temporal pulse sequence of the target analysis region, perform multi-level decomposition on the quantum spatio-temporal pulse sequence to obtain a decomposed pulse sequence, use the decomposed pulse sequence to construct a comprehensive correlation map of the dynamic scene, perform reinforcement learning on the comprehensive correlation map to obtain an optimized correlation map, and use the optimized correlation map to perform target recognition on the dynamic scene.

[0010] Optionally, the separating interference features from the quantum dynamic image based on the quantum polarization characteristics to obtain a clean information sequence includes:

[0011] Based on the quantum polarization characteristics, perform quantum substrate drift correction on the quantum dynamic image to obtain a polarization-corrected image;

[0012] Perform quantum polarization dimension projection on the polarization-corrected image to obtain projection features;

[0013] Perform adaptive quantum filtering on the projection features to obtain a filtered information sequence;

[0014] Calculate the information purity of the filtered information sequence;

[0015] Based on the information purity, perform benchmark polarization distribution reconstruction on the quantum dynamic image to obtain a clean information sequence.

[0016] Optionally, the performing optical-magnetic conversion on the clean information sequence using the optical signal encoding matrix to obtain a pulse feature map includes:

[0017] Perform element matching on the optical signal encoding matrix and the clean information sequence to determine the optical signal attributes of the clean information sequence;

[0018] Perform vector encoding on the optical signal attributes to obtain an optical signal vector;

[0019] Perform optical-magnetic conversion on the optical signal vector to obtain a magnetic signal vector;

[0020] Use a pre-configured mapping function to perform spatial mapping on the magnetic signal vector to obtain a pulse feature map.

[0021] Optionally, the identifying quantum correlation features of the pulse feature map includes:

[0022] Perform three-dimensional quantum grid partitioning on the pulse feature map to obtain quantum partitioning units;

[0023] Calculate the density matrix of the quantum partitioning unit;

[0024] Based on the density matrix, identify the quantum entanglement points of the pulse feature map;

[0025] Based on the quantum entanglement points, construct a quantum distribution histogram of the pulse feature map;

[0026] Based on the quantum distribution histogram, identify the quantum correlation features of the pulse feature map.

[0027] Optionally, the topological mapping of the fusion feature tensor to obtain a mapping feature includes:

[0028] Perform dimensionality reduction processing on the fusion feature tensor to obtain a dimensionality-reduced feature tensor;

[0029] Identify the topological structure characteristics of the dimensionality-reduced feature tensor;

[0030] Based on the topological structure characteristics, perform multi-level mapping on the dimensionality-reduced feature tensor to obtain a multi-level mapping tensor;

[0031] Extract the multi-level dimension features of the multi-level mapping tensor;

[0032] Perform adaptive scale adjustment on the multi-level dimension features to obtain adjusted features;

[0033] Perform encoding integration on the adjusted features to obtain mapping features.

[0034] Optionally, the calculation of the scene attention weight of the dynamic scene using the mapping feature includes:

[0035] Calculate the Pearson correlation coefficient matrix between different features in the mapping feature;

[0036] Based on the Pearson correlation coefficient matrix, construct an importance scoring function for the dynamic scene, where the importance scoring function can be represented in the following form:

[0037] where F represents the importance scoring function, represents the balance parameter of the importance scoring function, represents the i-th feature in the mapping feature, represents the j-th feature in the mapping feature, n represents the number of dimensions of the mapping feature, represents the k-th feature of the mapping feature, represents the Pearson correlation coefficient matrix between the i-th feature and the j-th feature in the mapping feature, represents the maximum query function of representation maximum query function;

[0038] Configure a spatial position weight function for the important score function to obtain a scene attention weight function;

[0039] Use the scene attention weight function to calculate the scene attention weight of the dynamic scene:

[0040] wherein, represents the scene attention weight, represents the important score of the i-th mapped feature, represents the spatial position of the i-th mapped feature, represents the important score of the j-th mapped feature, represents the spatial position of the j-th mapped feature, and n represents the dimension of the mapped feature.

[0041] Optionally, based on the scene attention weight, performing region separation on the quantum dynamic image to obtain a target analysis region, including:

[0042] Performing comprehensive feature analysis on the quantum dynamic image to obtain comprehensive quantum image features;

[0043] Construct a feature correlation matrix of the quantum dynamic image to analyze the feature correlation of the quantum dynamic image using the feature correlation matrix;

[0044] Based on the comprehensive quantum image features and the feature correlation, perform parameter correction on the scene attention weight to obtain an optimized weight;

[0045] Identify the region characteristics of the quantum dynamic image;

[0046] Based on the region characteristics, use the optimized weight to perform region separation on the quantum dynamic image to obtain a target analysis region.

[0047] Optionally, performing multi-level decomposition on the quantum spatio-temporal pulse sequence to obtain a decomposed pulse sequence, including:

[0048] Extract the information features of the quantum spatio-temporal pulse sequence;

[0049] Based on the information features, configure the wavelet basis function of the quantum spatio-temporal pulse sequence;

[0050] Use the wavelet basis function to calculate the detail coefficients and approximation coefficients of the quantum spatio-temporal pulse sequence at different scales;

[0051] Based on the detailed coefficient and the approximation coefficient, perform multi-level decomposition on the quantum spacetime pulse sequence to obtain a decomposed pulse sequence.

[0052] Optionally, the constructing the comprehensive association map of the dynamic scene by using the decomposed pulse sequence includes:

[0053] Perform quantization processing on the decomposed pulse sequence to obtain a quantization sequence;

[0054] Perform tensor representation on the quantization sequence to obtain a pulse tensor;

[0055] Extract features from the pulse tensor to obtain tensor features;

[0056] Based on the tensor features, construct the association network nodes of the dynamic scene;

[0057] Analyze the association relationships between the association network nodes to obtain association information;

[0058] Based on the association information, construct the initial network of the dynamic scene;

[0059] Define the cellular rules of the initial network;

[0060] Based on the cellular rules, perform iterative evolution on the initial network to obtain an evolved network;

[0061] Perform visual composition on the evolved network to obtain a comprehensive association map.

[0062] To solve the above problems, the present invention also provides a neuromorphic vision computing system for rapid recognition of dynamic scenes, and the system includes:

[0063] An interference feature separation module, configured to deploy a visual camera array in a dynamic scene to be rapidly recognized to obtain a quantum dynamic image of the dynamic scene, identify the quantum polarization characteristics of the quantum dynamic image, and based on the quantum polarization characteristics, perform interference feature separation on the quantum dynamic image to obtain a clean information sequence;

[0064] A feature fusion module, configured to construct an optical signal encoding matrix of the clean information sequence, use the optical signal encoding matrix to perform optical-magnetic conversion on the clean information sequence to obtain a pulse feature map, identify the quantum association characteristics of the pulse feature map, and based on the quantum association characteristics, perform quantum feature fusion on the pulse feature map to obtain a fusion feature tensor;

[0065] A target region separation module is used to perform topological mapping on the fused feature tensor to obtain mapping features, calculate the scene attention weight of the dynamic scene using the mapping features, and perform region separation on the quantum dynamic image based on the scene attention weight to obtain a target analysis region;

[0066] The target recognition module is used to collect the quantum space-time pulse sequence of the target analysis area, perform multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence, use the decomposed pulse sequence to construct a comprehensive association map of the dynamic scene, perform reinforcement learning on the comprehensive association map to obtain an optimized association map, and use the optimized association map to perform target recognition on the dynamic scene.

[0067] Compared with the problems described in the background technology, the present invention first deploys a visual camera array in a dynamic scene to be quickly identified, acquires quantum dynamic images, and identifies its quantum polarization characteristics, and uses quantum state measurement technology to acquire information such as polarization direction and intensity, so as to separate interference features, improve data quality, and reduce interference; further, the present invention constructs an optical signal coding matrix of a clean information sequence, and then uses wavelength division multiplexing technology to establish a mapping relationship between information and optical signal properties to enhance the signal's anti-interference ability, and performs optical-magnetic conversion on the clean information sequence, so that the information can be converted between optical and magnetic media, which is convenient for in-depth analysis of the scene; further, the present invention mines the hidden information of dynamic scenes through operations such as three-dimensional quantum grid division, calculation of density matrix, identification of quantum entanglement points, and construction of quantum distribution histograms, and based on quantum correlation characteristics, performs quantum feature fusion on the pulse feature map, and uses quantum entanglement gate operations and tensor network collection to obtain quantum information. The fused feature tensor is obtained by compression calculation to realize the integrated features, thereby improving the efficiency of image analysis; the present invention simplifies the data processing steps by performing dimensionality reduction processing on the fused feature tensor, identifying topological structure characteristics, multi-level mapping, extracting multi-layer dimensional features, adaptive scale adjustment and encoding integration to obtain mapping features, and then based on the scene attention weight, the quantum dynamic image is regionally separated to more deeply explore the detailed information of key areas in the dynamic scene, so as to have a more accurate understanding and cognition of the entire dynamic scene. For example, in applications such as target recognition, the target area that needs to be focused on can be quickly located; further, the present invention constructs a comprehensive correlation map of the dynamic scene and performs reinforcement learning on the comprehensive correlation map, so that the map can more accurately reflect the scene rules and target characteristics, and finally uses the optimized correlation map to identify the target of the dynamic scene, so as to realize flexible recognition of highly complex dynamic scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A schematic diagram of a flow chart of a neuromorphic visual computing method for rapid recognition of dynamic scenes provided by an embodiment of the present invention;

[0069] Figure 2 This is a schematic diagram of the modules of the neuromorphic vision computing system for realizing the fast recognition of dynamic scenes provided by an embodiment of the present invention.

[0070] The implementation, functional features, and advantages of the present invention will be further described in conjunction with embodiments and with reference to the accompanying drawings. Specific Embodiments

[0071] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0072] The embodiments of the present application provide a neuromorphic vision computing method for fast recognition of dynamic scenes. The execution subject of the neuromorphic vision computing method for fast recognition of dynamic scenes includes, but is not limited to, at least one of electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiments of the present application. In other words, the neuromorphic vision computing method for fast recognition of dynamic scenes can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to: a single server, a server cluster, a cloud server, or a cloud server cluster, etc.

[0073] Embodiment 1:

[0074] Referring to Figure 1 As shown, this is a flowchart of the neuromorphic vision computing method for fast recognition of dynamic scenes provided by an embodiment of the present invention. In this embodiment, the neuromorphic vision computing method for fast recognition of dynamic scenes includes:

[0075] S1. Deploy a visual camera array in the dynamic scene to be quickly recognized to obtain a quantum dynamic image of the dynamic scene, identify the quantum polarization characteristics of the quantum dynamic image, and based on the quantum polarization characteristics, separate interference features from the quantum dynamic image to obtain a clean information sequence.

[0076] In the embodiment of the present invention, by deploying a visual camera array in the dynamic scene to be quickly recognized to obtain a quantum dynamic image of the dynamic scene, the dynamic scene can be comprehensively photographed from different angles and positions to record the dynamic scene in the form of a quantum-related image.

[0077] Among them, the dynamic scene refers to a complex scene that combines object movement, environmental changes, and dynamic changes in quantum characteristics, such as an urban traffic intersection, an automobile manufacturing production line, and a drone reconnaissance scenario.

[0078] The visual camera array refers to a system composed of multiple visual cameras combined in a specific layout and manner. The quantum dynamic image refers to a special form of image that contains information such as the shape, color, and position of objects in the scene possessed by traditional images, and incorporates quantum-related characteristics. In a dynamic scene, the image changes over time, and the quantum dynamic image emphasizes the distribution and evolution of the quantum states of photons (such as quantum polarization, quantum entanglement, etc.) over time and space during this change process.

[0079] Furthermore, through identifying the quantum polarization characteristics of the quantum dynamic image in the embodiments of the present invention, information such as the change in the polarization direction and polarization state of light during propagation can be understood, facilitating the analysis of the specific change laws presented by the dynamic scene under different substances, environments, and events.

[0080] Among them, the quantum polarization characteristic refers to the unique properties exhibited by quantum in terms of polarization state, such as characteristics like quantum state superposition and quantum entanglement.

[0081] Optionally, the quantum polarization characteristic can be obtained by using quantum state measurement technology to directly measure the quantum polarization state in the quantum dynamic image and acquire information such as its polarization direction and polarization intensity.

[0082] Even further, in the embodiments of the present invention, by separating the interference characteristics of the quantum dynamic image based on the quantum polarization characteristic to obtain a clean information sequence, the data quality can be improved and the influence of interference information on subsequent analysis can be reduced.

[0083] As an embodiment of the present invention, separating the interference characteristics of the quantum dynamic image based on the quantum polarization characteristic to obtain a clean information sequence includes: performing quantum basis drift correction on the quantum dynamic image based on the quantum polarization characteristic to obtain a polarization correction image, performing quantum polarization dimension projection on the polarization correction image to obtain projection characteristics, performing adaptive quantum filtering on the projection characteristics to obtain a filtered information sequence, calculating the information purity of the filtered information sequence, and reconstructing the reference polarization distribution of the quantum dynamic image based on the information purity to obtain a clean information sequence.

[0084] Among them, the polarization correction image refers to the image obtained after performing quantum basis drift correction on the quantum dynamic image, and the information purity refers to an index used to measure the proportion of effective information in the filtered information sequence.

[0085] Optionally, the polarization-corrected image can be obtained by performing frame-by-frame analysis on the quantum dynamic image to determine the base state of quantum polarization in each frame of the image, then using quantum state measurement technology to detect the drift amount of the quantum base in each frame of the image in the time dimension, and then performing reverse correction on the quantum polarization data of each frame of the image according to the drift amount. The projection feature can be obtained by first identifying the dimensional space of the quantum polarization of the polarization-corrected image and then using the quantum projection algorithm to project the quantum polarization information of each pixel point in the polarization-corrected image into the dimensional space. The filtered information sequence can be obtained by first identifying the statistical characteristics of the projection feature, such as mean, variance, etc., and then using the least mean square error algorithm to filter the projection feature. The information purity can be obtained by using the von Neumann entropy formula and performing mathematical modeling on the quantum state of the filtered information sequence. The clean information sequence can be obtained by determining the parameters of the reference polarization distribution according to the information purity, such as the weight of the polarization direction, the polarization intensity, and then using an image reconstruction algorithm, such as a model-based reconstruction method, to adjust the image according to the reference polarization distribution model.

[0086] S2. Construct an optical signal encoding matrix for the clean information sequence, use the optical signal encoding matrix to perform optomagnetic conversion on the clean information sequence to obtain a pulse feature map, identify the quantum correlation feature of the pulse feature map, and based on the quantum correlation feature, perform quantum feature fusion on the pulse feature map to obtain a fusion feature tensor.

[0087] In the embodiment of the present invention, by constructing the optical signal encoding matrix for the clean information sequence, the information elements in the clean information sequence can be associated with specific attributes of the optical signal (such as wavelength, phase, amplitude, etc.), thereby facilitating the processing of the characteristics of the optical signal for the clean information sequence and improving the anti-interference ability of the signal.

[0088] Among them, the optical signal encoding matrix refers to a mathematical matrix used to convert the clean information sequence into an optical signal representation form, which defines the mapping relationship between each element in the clean information sequence and specific attributes of the optical signal (such as wavelength, phase, amplitude, polarization, etc.).

[0089] Optionally, the optical signal encoding matrix can use wavelength division multiplexing technology to select an appropriate number of wavelength channels according to the length and complexity of the clean information sequence. According to the length and complexity of the clean information sequence, select an appropriate number of wavelength channels. Divide the information sequence into multiple sub-sequences (each sub-sequence corresponds to a specific wavelength channel), set parameters such as the intensity and phase of the optical signal on each wavelength channel, and perform information encoding on the clean information sequence to obtain the optical signal encoding matrix.

[0090] The embodiment of the present invention utilizes the optical signal encoding matrix to perform optical-magnetic conversion on the clean information sequence to obtain a pulse characteristic diagram, which can enable information to be flexibly converted between two physical media, optical and magnetic, and is conducive to using magnetic information processing technology to deeply understand and analyze complex dynamic scenes.

[0091] The pulse characteristic diagram refers to an image with specific information generated by quantum processing of a dynamic scene.

[0092] As an embodiment of the present invention, the optical signal coding matrix is ​​used to perform optical-magnetic conversion on the clean information sequence to obtain a pulse characteristic diagram, including: element matching between the optical signal coding matrix and the clean information sequence to determine the optical signal properties of the clean information sequence, vector encoding the optical signal properties to obtain an optical signal vector, optical-magnetic conversion on the optical signal vector to obtain a magnetic signal vector, and spatially mapping the magnetic signal vector using a preconfigured mapping function to obtain a pulse characteristic diagram.

[0093] Among them, the optical signal vector refers to the vector obtained by quantizing and vectorizing the properties of the optical signal, the magnetic signal vector refers to the vector obtained after the optical signal vector is subjected to optical-magnetic conversion, and the preconfigured mapping function refers to a pre-set mathematical function used to map each component in the magnetic signal vector to the signal space, which can be configured using the feature extraction algorithm in the machine learning algorithm.

[0094] Optionally, the optical signal attribute can be obtained by accurately matching and locating each element in the clean information sequence according to the mapping relationship set by the coding matrix using a data matching algorithm. For example, if the coding matrix is ​​encoded in binary form, it can be determined based on the corresponding relationship of the binary bits. The optical signal vector can use vector construction technology to determine the vector dimension and component order according to the physical characteristics of the optical signal attributes, such as the wavelength, phase and other attributes of the light, and arrange them into a multidimensional vector (such as the length of the vector) in a certain order. The magnetic signal vector can be obtained by directly converting the optical signal vector using an optical-magnetic conversion device. The pulse characteristic diagram can use a spatial mapping algorithm and a preconfigured mapping function to map each component in the magnetic signal vector to the signal space to form a pulse distribution, and is constructed according to the pulse distribution.

[0095] Furthermore, the embodiments of the present invention are helpful in identifying important information such as the microstructure of objects in highly complex dynamic scenes, material interactions, etc. by identifying the quantum correlation characteristics of the pulse characteristic graph, thereby mining hidden information in the target scene.

[0096] Among them, the quantum correlation characteristics refer to a comprehensive description of the quantum-level relationships and characteristics in the pulse characteristic diagram, such as entanglement characteristics, correlation strength and distribution characteristics.

[0097] As an embodiment of the present invention, the identification of the quantum correlation features of the pulse feature map includes: performing a three-dimensional quantum grid division on the pulse feature map to obtain quantum division units, calculating the density matrix of the quantum division units, identifying the quantum entanglement points of the pulse feature map based on the density matrix, constructing a quantum distribution histogram of the pulse feature map based on the quantum entanglement points, and identifying the quantum correlation features of the pulse feature map based on the quantum distribution histogram.

[0098] Among them, the density matrix refers to a mathematical tool for describing the state of a quantum system. The quantum entanglement point refers to a concept in quantum mechanics that describes the special correlation state of a quantum system. For example, in the context of analyzing a pulse feature map, it refers to the points in the regionalized map obtained by regionalizing the pulse feature map that are determined to be closely related to quantum correlation through quantum measurement techniques. There is a quantum entanglement phenomenon between the quantum states represented by these points, that is, when one quantum state changes, the other quantum states entangled with it will instantaneously change accordingly, regardless of how far apart they are in space. The quantum distribution histogram refers to a chart used to visually display the distribution of quantum entanglement points in the pulse feature map.

[0099] Optionally, the quantum division unit can be obtained by constructing a three-dimensional grid framework based on the Heisenberg uncertainty principle according to the dimensions of the pulse feature map and the requirements of quantum characteristic analysis, and then determining the grid nodes at a specific spacing to divide the map into many small three-dimensional regions. The density matrix can be calculated using the compressive sensing tomography algorithm. The quantum entanglement points can be identified by using a quantum entanglement detection algorithm in combination with the eigenvalues and eigenvectors of the density matrix. The quantum distribution histogram can count the spatial distribution of quantum entanglement points in the pulse feature map, classify them according to different spatial positions or quantum entanglement intensity intervals, and then present the number distribution of quantum entanglement points in each interval in the form of a histogram. The quantum correlation features can be identified by observing the shape, peak position, and proportion of different intervals of the quantum distribution histogram.

[0100] Furthermore, in the embodiment of the present invention, by performing quantum feature fusion on the pulse feature map based on the quantum correlation features, a fusion feature tensor can be obtained, which can integrate different aspects of quantum features in the pulse feature map and combine quantum correlation features to form a more comprehensive and representative feature set to improve the image analysis efficiency.

[0101] Optionally, the fused feature tensor can encode the spatio-temporal information of the pulse feature map into a quantum state through quantum entanglement gate operations, and then perform tensor network contraction calculations on the pulse feature map of the quantum state in the Hilbert space. After optimizing the entanglement weights between the features in the pulse feature map after tensor network contraction using a variational quantum eigensolver, the fused feature tensor is obtained.

[0102] S3. Perform a topological mapping on the fused feature tensor to obtain a mapped feature. Use the mapped feature to calculate the scene attention weight of the dynamic scene. Based on the scene attention weight, perform region separation on the quantum dynamic image to obtain a target analysis region.

[0103] In an embodiment of the present invention, by performing a topological mapping on the fused feature tensor to obtain a mapped feature, the high-dimensional and complex fused feature tensor can be converted into a lower-dimensional feature that is easier to understand and process through topological mapping, simplifying the difficulty of data processing.

[0104] As an embodiment of the present invention, performing a topological mapping on the fused feature tensor to obtain a mapped feature includes: performing dimensionality reduction processing on the fused feature tensor to obtain a dimensionality-reduced feature tensor, identifying the topological structure characteristics of the dimensionality-reduced feature tensor, based on the topological structure characteristics, performing multi-level mapping on the dimensionality-reduced feature tensor to obtain a multi-level mapped tensor, extracting the multi-level dimensional features of the multi-level mapped tensor, performing adaptive scale adjustment on the multi-level dimensional features to obtain an adjusted feature, and performing encoding integration on the adjusted feature to obtain a mapped feature.

[0105] Among them, the topological structure characteristics refer to the description of the inherent properties of the dimensionality-reduced feature tensor at the topological level, such as critical point characteristics, connection relationships, and overall topological forms.

[0106] Optionally, the dimensionality-reduced feature tensor can be obtained by calculating the covariance matrix among the dimensions of the fused feature tensor using the principal component analysis algorithm, and then decomposing the eigenvalues of the covariance matrix to obtain the principal components. The topological structure characteristics can be determined by using Morse theory to calculate the gradient vector field of each data point in the dimensionality-reduced feature tensor, finding the critical points where the gradient is zero, and then determining the topological structure of the tensor based on the types of critical points (such as saddle points, maximum points, minimum points) and their connection relationships. The multi-level mapping tensor can be obtained by using the locally linear embedding algorithm to calculate the embedding coordinates of different dimensions of the dimensionality-reduced feature tensor in a low-dimensional space. The multi-level dimensional features can be obtained by calculating the geometric features (such as distance, angle), statistical features (such as mean, variance), and distribution features (such as clustering degree) of the multi-level mapping tensor. The adjusted features can be obtained by using an adaptive weight allocation algorithm to perform feature scaling on the multi-level dimensional features and extract key features (extracted according to the size of the weight allocation, and the weight can be set to 0.3, and the features with weights greater than 0.3 are retained, and specific settings need to be combined with the actual application). The mapping features can be obtained by using the Huffman coding method to encode the adjusted features and combine the features of different layers in a specific order.

[0107] Further, in the embodiment of the present invention, by using the mapping features to calculate the scene attention weight of the dynamic scene, the importance differences of different regions or features in the dynamic scene can be identified, which is beneficial to identifying the key information in the dynamic scene.

[0108] Wherein, the scene attention weight is a quantitative index used to indicate the relative importance degree of different regions or features in the dynamic scene.

[0109] As an embodiment of the present invention, the method of using the mapping features to calculate the scene attention weight of the dynamic scene includes: calculating the Pearson correlation coefficient matrix among different features in the mapping features, and constructing an importance scoring function for the dynamic scene based on the Pearson correlation coefficient matrix, where the importance scoring function can be expressed in the following form:

[0110] Where F represents the importance scoring function, represents the balance parameter of the importance scoring function, represents the i-th feature in the mapping features, represents the j-th feature in the mapping features, n represents the number of dimensions of the mapping features, represents the k-th feature of the mapping features, represents the Pearson correlation coefficient matrix between the i-th feature and the j-th feature in the mapping features, represents the maximum query function of represents The maximum query function;

[0111] Configure a spatial position weight function for the important score function to obtain a scene attention weight function, and use the scene attention weight function to calculate the scene attention weight of the dynamic scene:

[0112] Wherein, represents the scene attention weight, represents the important score of the i-th mapped feature, represents the spatial position of the i-th mapped feature, represents the important score of the j-th mapped feature, represents the spatial position of the j-th mapped feature, and n represents the dimension of the mapped feature.

[0113] Wherein, the Pearson correlation coefficient matrix refers to a matrix used to measure the degree and direction of linear correlation between different features in the mapped features. The important score function refers to a function that comprehensively considers the size of the mapped feature itself and its correlation with other features. The spatial position weight function refers to a mathematical formula used to measure the influence of the spatial position of the mapped feature in the dynamic scene on its importance. The balance parameter of the important score function refers to a parameter used to adjust the relative influence degree of the size of the feature itself and the correlation between features in the important score function.

[0114] Optionally, the Pearson correlation coefficient matrix can be calculated for the mapped features by using the Pearson correlation coefficient method. The spatial position weight function can be selected according to specific application scenarios and requirements. For example, in image or scene analysis, if it is considered that the features in the central region are more important and the importance gradually decreases in the regions far from the center, a Gaussian function can be used to configure the spatial position weight function. For example, when processing a picture, the central part of the image may contain the main target object, then a higher weight can be given to the central region through the Gaussian function, while the weight of the edge region is lower.

[0115] It should be further noted that the important score function comprehensively considers the size of the feature itself and the correlation between features to evaluate the importance of each feature in the dynamic scene, and adjusts the relative importance of these two factors through the balance parameter When is close to 1, it emphasizes the value of the feature itself; when is close to 0, it focuses on the correlation between features, and thus can more comprehensively and flexibly reflect the importance of different features in the dynamic scene. The value range of the balance parameter is [0 - 1].

[0116] Furthermore, in the embodiment of the present invention, by performing region separation on the quantum dynamic image based on the scene attention weight, the target analysis region can be obtained, which can concentrate the analysis resources on the most valuable part and improve the analysis efficiency. At the same time, it helps to more deeply explore the detailed information of the key regions in the dynamic scene, so that the understanding and cognition of the entire dynamic scene are more accurate. For example, in applications such as target recognition, it can quickly locate the region where the target to be focused on is located.

[0117] As an embodiment of the present invention, the performing region separation on the quantum dynamic image based on the scene attention weight to obtain the target analysis region includes: performing a comprehensive characteristic analysis on the quantum dynamic image to obtain the comprehensive quantum image characteristics, constructing a feature correlation matrix of the quantum dynamic image to analyze the feature correlation of the quantum dynamic image by using the feature correlation matrix, correcting the parameters of the scene attention weight based on the comprehensive quantum image characteristics and the feature correlation to obtain the optimized weight, identifying the region characteristics of the quantum dynamic image, and performing region separation on the quantum dynamic image by using the optimized weight based on the region characteristics to obtain the target analysis region.

[0118] Wherein, the comprehensive quantum image characteristics refer to a description set of the multi-faceted characteristics of the quantum dynamic image, the feature correlation matrix refers to a matrix used to quantify the correlation degree between different features in the quantum dynamic image, and the region characteristics refer to the relevant attributes of different regions in the quantum dynamic image, such as the type of quantum state, distribution uniformity and other attributes.

[0119] Optionally, the comprehensive quantum image features can be determined from dimensions such as quantum state distribution and frequency characteristics. Specifically, the quantum Fourier transform is used to transform the quantum dynamic image from the time domain to the frequency domain to obtain the frequency distribution characteristics of the image. Meanwhile, quantum state tomography technology is utilized to accurately measure the specific parameters of each quantum state in the image, such as amplitude, phase, etc. The feature correlation matrix can calculate key quantum features such as quantum entanglement points and qubits of the quantum dynamic image using quantum correlation functions, and then arrange the calculated correlation strength values according to the corresponding relationships of the features to construct a matrix. Then, by analyzing the matrix elements in the feature correlation matrix, such as calculating the eigenvalues and eigenvectors of the matrix, the feature correlation is identified. The optimization weights can be obtained by using the backpropagation algorithm in machine learning, taking the comprehensive quantum image features and feature correlation as inputs and the initial scene attention weights as parameters. By continuously adjusting the weight parameters, the output simulation region separation result of the algorithm is made to match the ideal image content distribution more closely. After multiple rounds of iterative training, the weights that highly match the actual features of the image are obtained. The regional characteristics can be obtained by using a quantum edge detection algorithm to utilize the sensitivity of quantum state changes to detect the edge contours of different regions in the image, and then combining with a quantum clustering algorithm, such as quantum K-means clustering, to cluster the pixel points in the image according to the similarity of quantum features, thereby determining characteristics such as the range, shape, and uniformity of internal quantum features of different regions. The target analysis region can apply the optimization weights to the image segmentation algorithm according to the identified regional characteristics. For regions with high weights, a fine quantum image segmentation method, such as segmentation based on quantum Monte Carlo simulation, is adopted to ensure the accurate division of key regions. Finally, the quantum dynamic image is divided into different regions, and the analysis regions with high weights are extracted, and regions with a weight value greater than 0.3 can be extracted.

[0120] S4. Collect the quantum space-time pulse sequence of the target analysis region, perform multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence, use the decomposed pulse sequence to construct a comprehensive correlation map of the dynamic scene, perform reinforcement learning on the comprehensive correlation map to obtain an optimized correlation map, and use the optimized correlation map to perform target recognition on the dynamic scene.

[0121] In the embodiment of the present invention, collecting the quantum space-time pulse sequence of the target analysis region can provide a raw data basis for subsequent in-depth analysis of the dynamic scene. These pulse sequences contain the change rules of quantum states in the space-time dimension, which helps to reveal the quantum characteristics of the targets in the dynamic scene and their evolution process over time, and plays a key role in accurately understanding and analyzing the dynamic scene.

[0122] Among them, the quantum space-time pulse sequence refers to a signal sequence that reflects the dynamic changes of a quantum system in the time and space dimensions.

[0123] Optionally, the quantum space-time pulse sequence can be collected by a quantum sensor array.

[0124] In the embodiment of the present invention, by performing multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence, the information mixed in the original pulse sequence can be separated, so that various features and patterns hidden therein can be revealed.

[0125] As an embodiment of the present invention, performing multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence includes: extracting the information features of the quantum space-time pulse sequence, configuring the wavelet basis function of the quantum space-time pulse sequence based on the information features, using the wavelet basis function to calculate the detail coefficients and approximation coefficients of the quantum space-time pulse sequence at different scales, and performing multi-level decomposition on the quantum space-time pulse sequence based on the detail coefficients and the approximation coefficients to obtain a decomposed pulse sequence.

[0126] Among them, the information feature refers to the key information extracted from the quantum space-time pulse sequence that can reflect its essential characteristics, such as energy distribution, phase, and amplitude, etc. The wavelet basis function refers to the basic function used for wavelet transform of the quantum space-time pulse sequence. The detail coefficient refers to the coefficient obtained when performing wavelet transform on the quantum space-time pulse sequence using the wavelet basis function, which reflects the local change characteristics of the pulse sequence in the high-frequency part. The approximation coefficient refers to the coefficient obtained when performing wavelet transform on the quantum space-time pulse sequence using the wavelet basis function, which reflects the overall trend of the quantum space-time pulse sequence in the low-frequency part.

[0127] Optionally, the information feature can be identified by performing time-domain conversion on the quantum space-time pulse sequence using quantum Fourier transform. The wavelet basis function can be configured by quantum Haar wavelet or Daubechies wavelet. For example, if there is important information in the low-frequency part of the pulse sequence, Haar wavelet can be used for configuration. The detail coefficients and the approximation coefficients can be obtained by performing inner product operations on the configured wavelet basis function and the quantum space-time pulse sequence at different scales. The decomposed pulse sequence can be obtained by using the quantum recursive algorithm starting from the approximation coefficient at the coarsest scale, decomposing the quantum space-time pulse sequence into the detail coefficients of the next layer and coarser approximation coefficients, continuously performing recursive operations, continuously decomposing the approximation coefficients of each layer, and obtaining the details and approximation coefficients at different scales, which are obtained by stacking these coefficients layer by layer.

[0128] Furthermore, in the embodiment of the present invention, by constructing a comprehensive correlation map of the dynamic scene using the decomposed pulse sequence, various information in the quantum space-time pulse sequence can be integrated together to form a visualizable or computable structure, which helps the computer to grasp the structure and evolution law of the dynamic scene as a whole.

[0129] Among them, the comprehensive association graph refers to a visualization graph used to present the complex relationships between various elements in a dynamic scenario.

[0130] As an embodiment of the present invention, constructing the comprehensive association graph of the dynamic scenario by using the decomposed pulse sequence includes: performing quantization processing on the decomposed pulse sequence to obtain a quantization sequence, performing tensor representation on the quantization sequence to obtain a pulse tensor, performing feature extraction on the pulse tensor to obtain tensor features, constructing the association network nodes of the dynamic scenario based on the tensor features, analyzing the association relationships between the association network nodes to obtain association information, constructing the initial network of the dynamic scenario based on the association information, defining the cellular rules of the initial network, iteratively evolving the initial network based on the cellular rules to obtain an evolved network, and performing visual composition on the evolved network to obtain the comprehensive association graph.

[0131] Among them, the cellular rule refers to the core concept in the cellular automaton model, which is a mechanism that defines how a cell updates and evolves according to its own state and the states of its surrounding neighbor cells.

[0132] Optionally, the quantization sequence can be obtained by using a non-uniform quantization algorithm to convert the continuous values of the decomposed pulse sequence into discrete values. The pulse tensor can be obtained by analyzing the information of the quantization sequence in dimensions such as time, space, and quantum characteristics to determine the order and dimensions of the tensor, and then arranging the data of the quantization sequence according to the information of each dimension. The tensor features can be obtained by using a tensor decomposition method, such as CP decomposition, to decompose the pulse tensor into multiple low-order tensors. The association network nodes can be determined by using a clustering algorithm, such as K-means clustering, according to the similarity and correlation of the tensor features, dividing the parts with similar features into a group, and each group corresponds to a node in the association network (the attributes of the node can be represented by the feature mean of the group), and sorting and screening the nodes by calculating the degree centrality of the nodes. The association information can be determined by using a quantum correlation analysis method to calculate the association strength between different association network nodes, and then analyzing the time sequence, spatial position relationship, etc. between the association network nodes to determine the association type of the nodes, and determining the association information based on the association strength and the association type. The initial network can be constructed based on the association network nodes and determining the connection relationships between the nodes according to the association information. Defining the cellular rules of the initial network can be achieved by determining the state update rules of the cellular automaton according to the physical characteristics and quantum laws of the dynamic scenario, and then setting the range and influence weights of the neighbor cells according to the influence mode and degree of the neighbor cells on the current cell. The comprehensive association graph can be obtained by using the Graphviz tool to draw according to the information of the nodes and edges of the evolved network.

[0133] Further, in the embodiments of the present invention, by performing reinforcement learning on the comprehensive association graph, an optimized association graph can be obtained, deficiencies and irrationalities in the comprehensive association graph can be discovered, and through iterative optimization, the graph can more accurately reflect the internal laws and target features of the dynamic scenario.

[0134] Optionally, to optimize the association graph, the state, action, and reward function of the comprehensive association graph can be defined first. The state covers information such as the graph structure and nodes, and actions include adjusting edge weights, etc. Then, using the reinforcement learning algorithm, select actions according to the state, execute to obtain a new state and reward, and continuously iterate and update the model parameters until an optimized association graph is obtained.

[0135] In the embodiments of the present invention, by using the optimized association graph to perform target recognition on the dynamic scenario, the target in the dynamic scenario can be more accurately recognized, especially in a complex and changing quantum dynamic scenario.

[0136] Embodiment 2:

[0137] As Figure 2 shown, it is a functional module diagram of the neuromorphic vision computing system for rapid recognition of dynamic scenarios of the present invention.

[0138] The neuromorphic vision computing system 200 for rapid recognition of dynamic scenarios of the present invention can be installed in an electronic device. According to the functions achieved, the neuromorphic vision computing system for rapid recognition of dynamic scenarios may include an interference feature separation module 201, a feature fusion module 202, a target area separation module 203, and a target recognition module 204. The modules in the present invention can also be referred to as units, which refer to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, and are stored in the memory of the electronic device.

[0139] In the embodiments of the present invention, the functions of each module / unit are as follows:

[0140] The interference feature separation module 201 is used to deploy a visual camera array in the dynamic scenario to be rapidly recognized, obtain the quantum dynamic image of the dynamic scenario, identify the quantum polarization characteristics of the quantum dynamic image, and based on the quantum polarization characteristics, perform interference feature separation on the quantum dynamic image to obtain a clean information sequence;

[0141] The feature fusion module 202 is used to construct an optical signal coding matrix of the clean information sequence, use the optical signal coding matrix to perform optical-magnetic conversion on the clean information sequence to obtain a pulse feature map, identify the quantum correlation characteristics of the pulse feature map, and based on the quantum correlation characteristics, perform quantum feature fusion on the pulse feature map to obtain a fusion feature tensor;

[0142] The target area separation module 203 is configured to perform a topological mapping on the fusion feature tensor to obtain a mapped feature, calculate a scene attention weight of the dynamic scene by using the mapped feature, and perform area separation on the quantum dynamic image based on the scene attention weight to obtain a target analysis area;

[0143] The target recognition module 204 is configured to collect a quantum space-time pulse sequence of the target analysis area, perform multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence, construct a comprehensive association map of the dynamic scene by using the decomposed pulse sequence, perform reinforcement learning on the comprehensive association map to obtain an optimized association map, and perform target recognition on the dynamic scene by using the optimized association map.

[0144] Specifically, each module in the neuromorphic vision computing system 200 for rapid recognition of a dynamic scene in the embodiments of the present invention adopts the same technical means as those in the above-mentioned Figure 1 neuromorphic vision computing method for rapid recognition of a dynamic scene, and can produce the same technical effects, which will not be elaborated here.

[0145] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A neuromorphic visual computing method for fast recognition of dynamic scenes, characterized in that: The method comprises: Deploy a visual camera array in a dynamic scene to be quickly identified to obtain a quantum dynamic image of the dynamic scene, identify the quantum polarization characteristics of the quantum dynamic image, and perform interference feature separation on the quantum dynamic image based on the quantum polarization characteristics to obtain a clean information sequence; Constructing an optical signal coding matrix of the clean information sequence, performing optical-magnetic conversion on the clean information sequence using the optical signal coding matrix to obtain a pulse characteristic diagram, identifying quantum correlation features of the pulse characteristic diagram, and performing quantum feature fusion on the pulse characteristic diagram based on the quantum correlation features to obtain a fused feature tensor; Performing topological mapping on the fused feature tensor to obtain mapping features, using the mapping features to calculate the scene attention weight of the dynamic scene, and performing regional separation on the quantum dynamic image based on the scene attention weight to obtain a target analysis area; The quantum space-time pulse sequence of the target analysis area is collected, and the quantum space-time pulse sequence is multi-level decomposed to obtain a decomposed pulse sequence, and the decomposed pulse sequence is used to construct a comprehensive association map of the dynamic scene, and reinforcement learning is performed on the comprehensive association map to obtain an optimized association map, and the optimized association map is used to perform target recognition for the dynamic scene.

2. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The step of performing interference feature separation on the quantum dynamic image based on the quantum polarization characteristics to obtain a clean information sequence includes: Based on the quantum polarization characteristics, performing quantum base drift correction on the quantum dynamic image to obtain a polarization corrected image; Performing quantum polarization dimension projection on the polarization-corrected image to obtain projection features; Performing adaptive quantum filtering on the projection features to obtain a filtering information sequence; Calculating the information purity of the filtered information sequence; Based on the information purity, the reference polarization distribution of the quantum dynamic image is reconstructed to obtain a clean information sequence.

3. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The method of using the optical signal encoding matrix to perform optical-magnetic conversion on the clean information sequence to obtain a pulse characteristic diagram includes: Performing element matching on the optical signal encoding matrix and the clean information sequence to determine an optical signal property of the clean information sequence; Performing vector encoding on the optical signal attribute to obtain an optical signal vector; Performing optical-magnetic conversion on the optical signal vector to obtain a magnetic signal vector; The magnetic signal vector is spatially mapped using a preconfigured mapping function to obtain a pulse characteristic diagram.

4. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The step of identifying the quantum correlation characteristics of the pulse characteristic graph comprises: Performing three-dimensional quantum grid division on the pulse characteristic diagram to obtain quantum division units; Calculating the density matrix of the quantum partition unit; Based on the density matrix, identifying quantum entanglement points of the pulse characteristic graph; Based on the quantum entanglement points, constructing a quantum distribution histogram of the pulse characteristic graph; Based on the quantum distribution histogram, a quantum correlation feature of the pulse signature is identified.

5. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The performing topological mapping on the fused feature tensor to obtain mapping features includes: Performing dimensionality reduction processing on the fused feature tensor to obtain a reduced-dimensional feature tensor; Identifying the topological structure characteristics of the reduced dimension feature tensor; Based on the topological structure characteristics, the dimension reduction feature tensor is subjected to multi-level mapping to obtain a multi-level mapping tensor; Extracting multi-layer dimensional features of the multi-level mapping tensor; Adaptively scaling the multi-layer dimensional features to obtain adjusted features; The adjustment features are encoded and integrated to obtain mapping features.

6. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The calculating the scene attention weight of the dynamic scene by using the mapping feature includes: Calculating a Pearson correlation coefficient matrix between different features in the mapping features; Based on the Pearson correlation coefficient matrix, an important score function of the dynamic scene is constructed, wherein the important score function is expressed in the following form: in, represents the important score function, represents the balance parameter of the importance score function, Indicates the mapping feature Features, Indicates the mapping feature Features, Indicates the number of dimensions of the mapping features, The first Features, Indicates the mapping feature Features and The Pearson correlation coefficient matrix of the features, express The maximum query function of express The maximum query function of ; Configuring a spatial position weight function for the importance score function to obtain a scene attention weight function; The scene attention weight function is used to calculate the scene attention weight of the dynamic scene: in, represents the scene attention weight, Indicates The importance scores of the mapped features, Indicates The spatial location of the mapped features, Indicates The importance scores of the mapped features, Indicates The spatial location of the mapped features, The number of dimensions representing the mapping features.

7. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The step of performing region separation on the quantum dynamic image based on the scene attention weight to obtain a target analysis region includes: Performing a comprehensive characteristic analysis on the quantum dynamic image to obtain comprehensive characteristics of the quantum image; Constructing a characteristic correlation matrix of the quantum dynamic image, so as to analyze the characteristic correlation of the quantum dynamic image by using the characteristic correlation matrix; Based on the comprehensive features of the quantum image and the feature correlation, the parameters of the scene attention weight are modified to obtain an optimized weight; identifying regional characteristics of the quantum dynamic image; Based on the regional characteristics and using the optimized weights, the quantum dynamic image is subjected to regional separation to obtain a target analysis region.

8. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The step of performing multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence comprises: Extracting information features of the quantum space-time pulse sequence; Based on the information characteristics, configuring the wavelet basis function of the quantum space-time pulse sequence; Utilizing the wavelet basis function, calculating detail coefficients and approximate coefficients of the quantum space-time pulse sequence at different scales; Based on the detail coefficients and the approximate coefficients, the quantum space-time pulse sequence is decomposed at multiple levels to obtain a decomposed pulse sequence.

9. The neuromorphic visual computing method for rapid recognition of dynamic scenes as claimed in claim 1, characterized in that: The method of constructing a comprehensive correlation map of the dynamic scene using the decomposed pulse sequence comprises: Quantizing the decomposed pulse sequence to obtain a quantized sequence; Performing tensor representation on the quantized sequence to obtain a pulse tensor; Performing feature extraction on the pulse tensor to obtain tensor features; Based on the tensor features, construct associated network nodes of the dynamic scene; Analyze the association relationship between the associated network nodes to obtain association information; Based on the association information, construct an initial network of the dynamic scene; Defining a cell rule of the initial network; Based on the cellular rule, iteratively evolve the initial network to obtain an evolved network; The evolutionary network is visualized to obtain a comprehensive correlation map.

10. A neuromorphic visual computing system for fast recognition of dynamic scenes, characterized in that: The system comprises: An interference feature separation module is used to deploy a visual camera array in a dynamic scene to be quickly identified to obtain a quantum dynamic image of the dynamic scene, identify the quantum polarization characteristics of the quantum dynamic image, and perform interference feature separation on the quantum dynamic image based on the quantum polarization characteristics to obtain a clean information sequence; A feature fusion module is used to construct an optical signal coding matrix of the clean information sequence, perform optical-magnetic conversion on the clean information sequence using the optical signal coding matrix to obtain a pulse feature diagram, identify quantum correlation features of the pulse feature diagram, and perform quantum feature fusion on the pulse feature diagram based on the quantum correlation features to obtain a fused feature tensor; A target region separation module is used to perform topological mapping on the fused feature tensor to obtain mapping features, calculate the scene attention weight of the dynamic scene using the mapping features, and perform region separation on the quantum dynamic image based on the scene attention weight to obtain a target analysis region; The target recognition module is used to collect the quantum space-time pulse sequence of the target analysis area, perform multi-level decomposition on the quantum space-time pulse sequence to obtain a decomposed pulse sequence, use the decomposed pulse sequence to construct a comprehensive association map of the dynamic scene, perform reinforcement learning on the comprehensive association map to obtain an optimized association map, and use the optimized association map to perform target recognition on the dynamic scene.

Images