183 results about "Visual interpretation" patented technology

The invention relates to the technical field of artificial intelligence and medical diagnosis, in particular to a disease diagnosis prediction method and system based on a graph neural network. The disease diagnosis prediction method based on the graph neural network comprises the five steps of heterogeneous medical knowledge graph construction, adaptive node embedding representation, hierarchical graph attention network modeling, incremental learning dynamic graph updating and multi-dimensional feature input and result output. The invention discloses a disease diagnosis and prediction system based on a graph neural network. The system comprises a multi-source data acquisition module, a heterogeneous graph construction module, a self-adaptive embedding module, a graph network calculation engine, a dynamic updating module, a disease prediction module and a feedback optimization module. According to the method, the multi-modal heterogeneous knowledge graph is constructed to integrate the multi-dimensional data of the patient, and the hierarchical graph attention network and the dynamic incremental learning are combined, so that the accurate prediction of the disease risk and the visual explanation of the pathological association path are realized.

Techniques are disclosed for enabling creators to create multimodal chatbots that are based on or simulate / model characters. The characters may be from audiovisual (AV) media such as films and TV shows or real people. The application leverages a combination of Visual Interpretation AI, Retrieval-Augmented Generation (RAG), Low-Rank Adaptation (LoRA), and function calling to provide rich, interactive experiences. The inferencing performed by an instant multimodal chatbot utilizes base weights, character weights, relationship weights, experience weights as we as environmental inputs. An instant chatbot uses a number of AI models including a video to text model, an image to text model, a sensory to text model, a large language model, a text to video model, a text to image model and a text to voice model. A user can interact with the chatbot in a variety of ways including text, audio and video.

The invention discloses a variable working condition migration bearing fault diagnosis method based on a ViTCapsNet network in a noise environment, and the method comprises the steps: firstly obtaining vibration signal data of a rolling bearing in different fault states under different loads, and adding Gaussian white noises of different degrees; performing continuous wavelet transform on the data added with the noise to obtain a two-dimensional wavelet image, and dividing the two-dimensional wavelet image into a training set and a test set; a ViT Encoder model is established, position information capable of being trained is added by using an Embedded module, position features are embedded into the model, and global information in the features is extracted through a multi-head self-attention mechanism in the Encoder module; constructing a capsule network which comprises a primary capsule layer and a digital capsule layer, further extracting spatial information in the features, and performing accurate fault diagnosis; and finally, performing visual interpretation on the features processed by the Encoder module by using Grad-CAM, and completing efficient migration fault diagnosis on a test data set.

The invention provides a talent cultivation recommendation method based on federal learning and natural language processing, and the method comprises the following steps: processing multi-source heterogeneous data, and obtaining standard data; constructing a talent knowledge graph capable of performing incremental updating; performing anonymization and desensitization processing on the sensitive data, and performing joint modeling by using a federated learning framework on the premise of not sharing original data to obtain a local model; talent portraits and post portraits are generated, and a talent matching model is obtained in combination with multiple target factors and a dynamic weight preset matching algorithm; continuously monitoring the similarity difference between the talent portrait and the post portrait, and dynamically generating a course optimization scheme and a post optimization scheme; and performing visual explanation or keyword description on the recommendation process. According to the method, matching precision and efficiency can be improved, data islands can be broken, privacy security can be protected, market demand changes can be dynamically adapted, personalized demands of multiple parties can be met, and accurate connection between education and employment can be promoted.

The invention belongs to the technical field of remote sensing semantic segmentation information extraction, and particularly relates to a remote sensing image land classification method based on SAM multi-order fine tuning. According to the method, remote sensing multi-classification precision and generalization ability can be improved. Firstly, a sample data set is constructed through visual interpretation based on domestic high-resolution series images; on the basis, a network is designed by utilizing the strong feature extraction capability of an SAM encoder, a domain offset adaptation module and a deep low-rank adaptation module are designed, a mask generation part of an original SAM is improved, and three-stage adaptation fine adjustment is performed on the SAM, so that the feature extraction capability of the network on a remote sensing image is enhanced, and the model performance is improved. In application, the remote sensing land utilization coverage multi-classification technology with practical value is obtained so as to practically promote application and development of a large model in remote sensing image semantic segmentation multi-classification.

The invention relates to a small sample fault prediction method based on physical information guidance and multi-source adaptive fusion. The method comprises the following steps: collecting real operation data of preprocessing target equipment; according to the physical model or domain knowledge of the target equipment, generating simulation sensor data conforming to a physical rule under various fault modes of different degrees; constructing diversified training samples in combination with real data and simulation data; for different types of sensor data, designing corresponding feature extraction branches, mining potential fault features in the data, and dynamically adjusting the weight of each data source fault feature for fusion based on an output result of a physical model and data-driven feature correlation analysis; inputting the obtained fusion features into a fault prediction model based on a small sample learning framework for training; comprehensively considering the fault prediction result, the real operation data and the analysis result of the physical model, and carrying out quantitative evaluation on the overall health state of the target equipment; and causal diagnosis and visual interpretation are carried out.

The invention relates to the technical field of coastline extraction, and particularly discloses a coastal intertidal zone waterline remote sensing extraction technology based on tide level data, which comprises the following steps: S1, acquiring a remote sensing image based on an earthplorer and a copernicius website and preprocessing the remote sensing image; s2, extracting an instantaneous waterline of the remote sensing image by using an NDWI index and a Canny operator to obtain an extraction result; s3, according to an extraction result, discretizing the obtained waterline, and by establishing a function relationship between a discrete point location and a tide level, obtaining an average large tide line so as to determine a preliminary coastline; and S4, correcting the coastline through visual interpretation to obtain a final coastline, and carrying out precision test. According to the coastline extraction method, the NDWI index and the Canny algorithm are used, visual interpretation and an automatic extraction method are combined, the waterline is discretized to determine the position of the average large tide line, and finally the more accurate and efficient coastline extraction method is obtained.

The invention discloses a data-based XGBoost model construction method for evaluating the cerebral hemorrhage recurrence risk of a patient, and the method comprises the steps: S10, reading and processing the data of the cerebral hemorrhage patient, and obtaining the data of a training set; s20, receiving the training set data, processing the training set data by adopting LASSO regression and a step-by-step recursion program, and outputting a plurality of variable features; s30, judging whether the p value of each characteristic variable is less than 0.05 or not; s40, when the p value of the characteristic variable is smaller than 0.05, outputting the variable characteristic; s50, constructing and optimizing parameters of the first model, and outputting an XGBoost model; and S60, carrying out visual interpretation on an evaluation result of the XGBoost model by using an SHAP (Secure Hardware Application Program) and an LIME (Local Instrumentation Model Element) interpreter Compared with the prior art, the method provided by the invention has the advantages that the model resolution AUC is 0.92 and the prediction accuracy is 0.95 by taking nine characteristic factors as internal verification of the model; the model resolution ratio AUC tested by the test set is 0.93, and the prediction accuracy is 0.91.

The invention belongs to the field of space earth observation technology and application, and discloses a remote sensing image stony desertification grading method, device and equipment based on a feature space, and the method comprises the steps: obtaining a remote sensing image, elevation data DEM and land utilization data of a research area, and carrying out the preprocessing; terrain correction is carried out on the area with the gradient larger than 25 degrees through an SCS + C terrain correction model; constructing a rock index RCRI and a vegetation index NDRE, and performing normalization; an RCRI-NDRE feature space is constructed; constructing a stony desertification index according to a distance formula from a point to a straight line; high-resolution remote sensing images are used for assisting visual interpretation of stony desertification sample points of different degrees, and grading thresholds between stony desertification of different degrees in the RCRI-NDRE feature space are calculated; and stony desertification grading is carried out based on the grading threshold value, and distribution conditions of stony desertification of different degrees are obtained. According to the method, remote sensing image data are extracted, and the RCRI-NDRE feature space of the rock index RCRI and the vegetation index NDRE is constructed, so that different degrees of stony desertification grading of the target area can be realized.

The invention discloses a sea ice type identification method and device based on a synthetic aperture radar image, and belongs to the technical field of satellite remote sensing marine monitoring. The method comprises the steps of 1, preprocessing a dual-polarized SAR image, and extracting input feature data; step 2, constructing a finely labeled sea ice type data set covering a generic north pole scene by combining visual interpretation and external auxiliary data; step 3, inputting the constructed finely labeled sea ice type data set into a sea ice type identification model improved based on a DeepLabv3 + architecture, and performing iterative optimization through a multi-graphics card parallel training mode to generate a high-precision sea ice type identification model; and 4, performing sea ice type identification on the extracted input feature data by using a high-precision sea ice type identification model. According to the method, the incidence angle dependency related to the sea ice type in the original SAR data is more accurately reserved, and the precision of the model is improved.

The invention belongs to the technical field of terrain remote sensing monitoring, and particularly relates to a surface deformation monitoring method based on an intermittent stacking time sequence InSAR (Intermittent Synthetic Aperture Radar), which comprises the following steps of: firstly, acquiring a plurality of SAR images, high-precision terrain data and orbit data of a target area in a lifting orbit within a period of time; performing geographical registration, interference combination selection, space-time filtering and phase unwrapping on the SAR image to generate a time sequence interferogram; calculating an annual average deformation rate of each pixel point in the interferogram, and generating an annual average deformation rate result of the target area; and finally, generating a suspicious landslide polygon according to an annual average deformation rate result, and performing visual interpretation and field investigation to obtain a real area of the landslide polygon. According to the method, result missing caused by the intermittent coherence characteristic of the pixel points can be avoided, the pixel selection has higher latitude, higher coverage can be achieved, the requirement for computing power is remarkably reduced, and a deformation rate result in a low-coherence environment can be rapidly obtained.

The invention belongs to but not limited to the technical field of animal husbandry statistics, and particularly relates to an alpine meadow yak number calculation method and system based on an unmanned aerial vehicle image, and the method comprises the steps: S1, employing an unmanned aerial vehicle to shoot an orthoimage of a sample plot, and carrying out the statistics of the yak number; s2, splicing the shot images by using Pix4D software, screening a proper sample plot, and respectively counting the quantity of wet cow dung and the quantity of dry cow dung in a visual interpretation mode; and S3, carrying out linear fitting on the obtained data by using Origin 2024b, so as to obtain three relation curves. According to the method for estimating the number of the yaks through statistics of the number of the cow dung in the unmanned aerial vehicle image, the workload of data collection is greatly reduced, and the limitation that the number of the yaks cannot be obtained when the yaks are not grazed in the grassland is avoided. According to the invention, the data acquisition frequency of experimenters is effectively reduced, and the research cost is reduced.

The invention discloses a water area and resource quantity monitoring and predicting method and system based on a remote sensing technology, and belongs to the technical field of water resource monitoring. The method comprises the following steps: determining a water area range according to historical data of a research area and an on-site survey condition, and determining a research area range; the method comprises the following steps: respectively acquiring ground data and water depth and water surface elevation data by using an unmanned aerial vehicle and an unmanned ship, respectively processing to obtain a land DEM and a water area DEM, and fusing to form an all-terrain DEM; downloading satellite image data of the research area, and performing geographical registration on the satellite image data; in the geographically registered satellite image data, describing the boundary of the water area through visual interpretation; superposing the water area boundary to an all-terrain DEM, obtaining the elevation of each point on the water area boundary line, calculating the average elevation of the water area boundary, and calculating the water area and the water resource amount to obtain the historical water area and the water resource amount; and on the basis of the obtained historical water area and water resource quantity, utilizing an ARIMA model to predict future water resource quantity and water area.

The invention discloses a landslide susceptible area grading evaluation method based on an extreme rainfall Thiessen polygon. Extreme rainfall is the most direct factor for inducing landslides, and the possibility of landslides in corresponding areas can be increased by increasing the extreme rainfall frequency and the extreme rainfall intensity. According to the method, a research area is divided into a plurality of Thiessen polygon areas, and only one rainfall station or weather station exists in each Thiessen polygon area; obtaining long-sequence day-by-day rainfall data, respectively calculating an extreme rainfall frequency index and an extreme rainfall intensity index of each station, and obtaining an extreme rainfall evaluation result of each Thiessen polygon region; performing visual interpretation or artificial intelligence interpretation on the remote sensing satellite image to obtain a large-range landslide layer; and superposing the extreme rainfall evaluation result of each Thiessen polygon region and the large-range remote sensing landslide interpretation result, and carrying out grading evaluation on the landslide-prone region. According to the invention, grading evaluation is carried out on the landslide-prone areas, and the prevention and control efficiency of landslide disasters is improved.

The invention relates to the technical field of literature measurement, and discloses a literature visual analysis method and system assisted by a large language model, and the method comprises the steps: matching a chart and an introduction document corresponding to the chart in a visual knowledge base according to an input problem; combining the input question and the introduction document to generate a first cue word, and inputting the first cue word into a large language model; the big language model outputs a structured language and a visual image reconstructed by a four-step method; interpreting the visual image by the visual large language model, and outputting visual interpretation data; visual interpretation and question input are combined to generate a third cue word, the third cue word is input into the large language model, and the large language model generates a final visual chart and chart visual interpretation; based on a large language model, by combining a customized knowledge base and a general knowledge base, the visualization diversity is increased, and high professional performance and strong generalization ability of unconventional problems are shown in the field of literature metering; and by adopting abnormal feedback, closed-loop control of the intelligent agent on a visual result is realized, and the reliability is remarkably improved.

The invention provides a forest degeneration process identification and degeneration degree division method based on a time sequence, and relates to the field of forest resource degeneration restoration. The method comprises the following steps: acquiring surface reflectance data, and calculating an NBR index after preprocessing; fitting an NBR time sequence through a LandTrendr algorithm, and extracting sample place time sequence data based on random sampling and visual interpretation; generating an adversarial network and expanding sample point fitting data into a simulation data set in combination with a self-supervised learning technology; training CNN, random forest and BOSSVS time sequence classification models, and constructing a hybrid classification model through an integrated voting strategy; and identifying a forest degeneration process by using a hybrid classification model, and generating a degeneration degree diagram through reclassification. The method can achieve the efficient recognition of the complex degradation process of the forest region, has good time sequence adaptability and regional applicability, and provides technical support for the monitoring and management of ecological restoration of regional forests.

The invention belongs to the field of laser processing, and relates to a method for reducing wall damage by regional light control based on a penetration model, which comprises the following steps of: 1) constructing the penetration model; 2) interpreting the penetration model constructed in the step 1) based on a camera visual interpretation technology to obtain an interpretation result; 3) capturing an image of an actual production process in real time, comparing the image of the actual production process with the interpretation result obtained in the step 2), and sending a control signal to a light control system according to a comparison result; and (4) the light control system adjusts the distribution of the light source in time according to the control signal in the step (3), and the actual drilling process is completed. The invention provides the method for reducing the damage to the wall by regional light control based on the penetration model, which can make a judgment in time and can reduce the strict requirement on the real-time performance of the system.

The invention relates to the technical field of carbon reserve estimation, and discloses a mangrove forest biomass carbon reserve estimation method based on a different-speed growth inversion model, and the method comprises the following steps: 1, setting a quadrat region; 2, collecting aerial images, kandelia candel morphological data, kandelia candel destructive data and soil profile organic carbon density data; step 3, establishing an allotropical growth inversion model; processing the kandelia candel destructive sampling parameters to obtain a kandelia candel carbon conversion coefficient; performing image processing and visual interpretation on the aerial image of the research area to obtain the crown breadth coverage total area and the mud flat coverage area of the kandelia candel in the research area; 4, the overall kandelia candel carbon reserve is obtained by combining the kandelia candel morphological data, the total crown breadth coverage area of the kandelia candel, the different-speed growth inversion model and the kandelia candel carbon conversion coefficient; and combining the organic carbon density data of the soil profile and the mud flat coverage area to obtain the overall soil carbon reserve. Meanwhile, the invention further provides a mangrove forest biomass carbon reserve estimation system based on the different-speed growth inversion model.

The invention discloses a remote sensing crop drawing-oriented systematic classification sample purification and expansion method. The method comprises the steps of open source remote sensing crop classification product data set acquisition, data preprocessing and availability condition analysis, sample purification based on a multi-source data space-time filtering algorithm, and sample expansion in combination with remote sensing feature enhancement and self-supervised deep learning. According to the method, high-quality crop samples which are high in representativeness, sufficient in number and balanced in space and category distribution can be efficiently obtained on the basis of multi-source crop classification product data, sample data support is provided for remote sensing crop classification research and application, and compared with traditional crop sample obtaining methods such as field collection and visual interpretation, the method has the advantage that the method is more efficient. The method is more efficient, objective and automatic, compared with an existing sample generation method based on simple processing of remote sensing crop classification product data, the method is more comprehensive, systematic and universal, and the generated samples have more advantages in the aspects of accuracy, representativeness and balance.

The invention discloses a judgment method suitable for a multi-branch river deformed river bay, and belongs to the field of water conservancy projects. The method comprises the steps of coarse extraction of a multi-branch river water body, depiction of a river multi-branch channel network structure, batch calculation of plane geometric indexes of all branch channels of the river, identification of main flow paths of the multi-branch river, and judgment and output of the multi-branch river abnormal river bay under the variable spatial scale. A traditional deformed river bay quantification and discrimination method is popularized to a multi-branch river with a complex plane form. The judgment method provided by the invention not only embodies the advantages of the remote sensing and image processing technology in the aspects of automation and calculation efficiency, but also considers the calculation precision close to that of visual interpretation, so that a large amount of cost is saved, and meanwhile, calculation errors caused by a data source and subjectivity are also avoided.

The embodiment of the invention discloses a causal inference method for traffic accident data processing and prediction model optimization, and the method comprises the steps: constructing a multi-source heterogeneous accident database, and constructing a standardized feature matrix; based on the standardized feature matrix, extracting a feature constraint rule through a decision tree, and generating a minority class anti-fact sample conforming to domain knowledge in combination with a heuristic algorithm to form an equalized data set; inputting the equalized data set into a double-path model system for training, expanding a feature space by a data path to improve the characterization capability of a minority class, and embedding an interpretability constraint module into an interpretation path to obtain a trained model; based on the trained model, constructing a visual interpretation system comprising a feature importance map and a causal association network; and based on a visual interpretation system and the standardized feature matrix, constructing a multi-level causal network identification core influence factor, and quantifying the causal effect of intervention measures through anti-fact simulation to form a priority intervention list.

The invention discloses an active visual explanation system for blind person field travel, and belongs to the technical field of man-machine interaction. Based on the system, a user can selectively broadcast navigation obstacle avoidance or scenery description, real-time image information collected by a deep vision detection module is recognized through a deep neural network model, speed and type information of objects in front and below is obtained in combination with a positioning system, and when the user selects to broadcast navigation obstacle avoidance, the navigation obstacle avoidance or scenery description is broadcasted. After the audio instruction sequence corresponding to each sub-state is generated according to the shortest state sequence, the user acts under the guidance of the audio instruction, and each action can be fed back by the system, so that the correctness and efficiency of the traveling forward route operation are greatly improved; when the user selects to broadcast the beautiful scenery description, the AI model generates an audio instruction sequence according to the identified object information and informs the user of the audio instruction sequence; not only is the safety of the user during the on-site travel improved, but also the user can be helped to know the scenic spot in which the user participates, and the whole-process visual supervision is provided for the on-site travel of the user.

The invention relates to a dual ERA-LFD kit for visually and rapidly detecting salmonella typhimurium or a single-phase variant thereof, a primer group and application of the dual ERA-LFD kit and the primer group, and aims at solving the problems that an existing detection method is complex in operation, long in time consumption and depends on professional equipment, and primer cross interference, high false positive rate and incapability of distinguishing the single-phase variant caused by the fact that dual ERA primer design is not standard. Screening to obtain a specific primer probe combination aiming at the fljB gene deletion region and the STM4495 gene; a dual ERA-LFD detection system is constructed, and double targets are synchronously amplified under a constant temperature condition; visual interpretation is realized through a double-detection-line color development mode of the transverse flow test strip. The detection specificity can be remarkably improved, and false positive is effectively avoided; the accurate distinguishing between the salmonella typhimurium and the single-phase variants is realized; the detection process is rapid and simple, does not need professional instruments and personnel, and is suitable for on-site rapid screening; the detection cost is obviously reduced and the basic application feasibility is improved.

The invention discloses a medical image segmentation method based on a Kolmogorov-Arnold network, and the method comprises the steps: firstly obtaining a medical image, then constructing an IKANet model, and carrying out the segmentation of the medical image after training, testing and verification. Wherein the IKANet model adopts an encoder-decoder structure and integrates three core components, namely a DKC module, a multi-scale attention module and an interpretability analysis module, and the DKC module realizes adaptive nonlinear transformation and feature decoupling through a Kolmogorov-Arnold network; the multi-scale attention module enhances space-channel feature fusion through multi-scale convolution and a channel attention mechanism; and the interpretability analysis module generates a multi-scale class activation mapping thermodynamic diagram and provides visual interpretation. The technical problems that an existing medical image segmentation method is limited in generalization ability and insufficient in interpretability are solved, and the method can be widely applied to data sets of colon polyp, pneumonia, pneumothorax and the like.

The invention discloses a disease control big data analysis method and system, and the method comprises the following steps: obtaining a data set X comprising a plurality of class attributes, and carrying out the preprocessing; performing independent optimal sequence search on each category attribute in the data set to obtain an optimal sequence, determining a distance structure suitable for disease control big data analysis, calculating distance measurement between a sample and a cluster according to the obtained optimal sequence, and constructing a clustering model; and by minimizing a target function, updating a sample attribution matrix and sequence, and dynamically adjusting a clustering result. Compared with the prior art, the method has the advantages that the system can adaptively adjust the clustering result by dynamically learning the optimal value sequence of the data and the corresponding distance structure, so that the clustering accuracy is remarkably improved, and visual interpretation is provided for understanding of the data.

The invention provides a nuclear magnetic image-based endometrial cancer classification and diagnosis system, and the working principle is that the advantages of artificial intelligence in image recognition and visual interpretation are utilized, a deep learning theory is adopted, a picture sample generated by clinical nuclear magnetic resonance examination is taken as a data basis, and image data are sliced, labeled and transformed to obtain an endometrial cancer classification and diagnosis result. A region of interest is selected from the data, and stage and classification diagnosis of the endometrial cancer is realized through a deep learning neural network model; therefore, the method does not need to perform an operation on the patient, can perform stage diagnosis on the endometrial cancer of the patient in a non-invasive manner, has the advantages of being non-invasive, accurate and the like, and can assist a doctor to perform professional and accurate diagnosis on the endometrial cancer in a classified manner; meanwhile, compared with common doctors, the method has higher diagnosis accuracy and higher diagnosis speed, can assist primary doctors in providing diagnosis opinions, and has important significance on treatment of endometrial cancer, improvement of the survival rate of patients and reduction of public health burden.

The invention discloses a TBM tunneling geological condition intelligent identification method and system fused with multiple modes. The method and system are used for identifying various geological condition scenes faced by a shield tunneling machine TBM during construction. The method comprises the steps that muck images and parameter information data are collected based on a shield tunneling machine industrial camera system and a sensor network, and data preprocessing operation is carried out; based on the collected data information, an improved Swin Transform model is provided, and multi-modal input of muck images and TBM parameters is fused; an improved multi-modal fusion model Swinin-T-M is integrated on the basis of a Dempster Shafer theory, the limitation and instability of a single model are improved, an integrated multi-modal deep learning model ESwinin-T-M is built, and the recognition accuracy and robustness of an algorithm are improved; and a visual interpretation technology is utilized to analyze the learning rule and key features of the model. According to the method, intelligent and accurate recognition of complex geological conditions in the TBM tunneling working process is achieved, and a feasible scheme is provided for detection of the construction environment of the shield tunneling machine.

The invention discloses an intelligent automatic drawing system and method, and relates to the technical field of agricultural remote sensing monitoring. The method comprises the following steps: acquiring a multi-temporal remote sensing image of a farmland area, performing preprocessing and spatial alignment, dividing the farmland into spatial units, extracting time sequence spectrum and vegetation index characteristics of the spatial units, and forming a visual characteristic vector sequence; analyzing the sequence by using unsupervised clustering, and automatically identifying the time sequence evolution trajectory categories representing different growth states; and finally, the system automatically marks the space units which are judged to be abnormal categories on a map, generates early warning and outputs a farmland abnormity thematic map. The system comprises a data preprocessing module, a data preprocessing module, a track classification module and an abnormal graph generation module. According to the invention, full-automatic and intelligent dynamic monitoring and abnormity identification of the farmland growth state are realized, the limitation of traditional single-temporal analysis or artificial visual interpretation is overcome, and the monitoring efficiency and objectivity are significantly improved.

The invention provides a method and a device for evaluating a credible result of a graph neural network. The method comprises the following steps: constructing an objective function based on a space-time relationship; obtaining an identification result obtained by identifying the skeleton behavior by the to-be-evaluated space-time diagram neural network, and constructing an augmented Lagrange function corresponding to the target function; iteratively calculating a node feature mask matrix and a human skeleton motion rotation model; and taking the node feature mask matrix and the human skeleton motion rotation model as an evaluation result of a prediction result of the space-time diagram neural network. According to the method, the to-be-evaluated space-time diagram neural network is explained and evaluated based on the human skeleton motion rotation model and the node feature mask corresponding to the human body. The interpretation result obtained by the method has high sparsity, and the geometric consistency of the space-time structure is consistent with that of the space-time diagram input into the space-time diagram neural network to be evaluated; the decision result of the space-time diagram network can be quantitatively and visually explained.

The invention discloses a reconstruction task interpretability analysis method based on an encoder-decoder architecture, and the method is called as Encoder-Decoder Recstructionality-Oriented Feature Activity Mapping (EDRAM), and comprises the following steps: loading a pre-training model, constructing an EDRAM class, carrying out image preprocessing, carrying out forward propagation on the model, calculating the EDRAM, fusing an EDRAM result, and carrying out visual analysis. The method is suitable for multi-category reconstruction tasks, and can provide independent interpretability analysis for each category and network hierarchy. Through fusion and visualization of multi-level EDRAM results, the EDRAM method realizes comprehensive and visual interpretation of the encoder-decoder model decision process, and provides an important basis for model optimization and performance improvement.