285 results about "Manual annotation" patented technology

The invention discloses a multi-granularity visual reasoning model construction method and device based on reinforcement learning. The method comprises the following steps: constructing an'image-reasoning query-bounding box 'triple as a training data set; designing a composite reward function including positioning precision, target counting precision and format reward; training the multi-modal large language model by adopting a GRPO algorithm; the trained model can output a region-level bounding box, and a pixel-level mask is generated and a contour-level result is extracted in combination with the segmentation model. According to the method, the problems that in the prior art, a reasoning path depends on manual annotation, multi-granularity tasks cannot be expanded, and generalization is insufficient are solved, and the autonomous decision-making ability, task expansibility and generalization in a distribution offset scene of the model are improved.

The invention discloses an abnormal data monitoring method and device based on artificial intelligence, and the method comprises the steps: 1, dividing an original data stream through a sliding window, extracting statistics, time sequence and change rate features, and dynamically screening features adaptive to data distribution based on an SHAP value; 2, constructing a double-flow model, capturing a global isolated mode by adopting an improved isolated forest in a static flow, capturing time sequence dependence on the basis of LSTM-AE in a dynamic flow, and fusing two-flow scores through performance-driven dynamic weight distribution; 3, combining a density peak value algorithm with historical density attenuation weighting, and dynamically adjusting an abnormal threshold value; 4, realizing low-delay incremental learning through a double-trigger mechanism and experience playback; 5, multi-granularity interpretation is generated, manual annotation feedback is supported, feature engineering and model training are integrated, and a'detection-interpretation-feedback-optimization 'closed loop is formed; high-adaptability anomaly monitoring is realized through dynamic feature screening, double-flow fusion detection, threshold value self-adaption and man-machine collaborative optimization.

The invention relates to the technical field of code development, and provides an AI code effective proportion statistical method and device, a medium and equipment. The method comprises the steps of obtaining related information of codes submitted by a user; according to a user name in the related information, searching log information of an AI code generated by a corresponding user through adoption of a code generation tool; under the condition that the file name of the code submitted by the user is matched with the file name in the log information, searching a corresponding submitted code segment from the code submitted by the user according to the mark information in the log information; and calculating the similarity between the AI code and the submitted code segment, and counting the effective proportion corresponding to the AI code according to the similarity. Therefore, the calculation of the effective proportion considers the quality of the AI code, so that the effective proportion can accurately reflect the real contribution of the AI code, and the calculation process is automatically realized, thereby avoiding the tedious, time-consuming and labor-consuming conditions of manual labeling.

The invention relates to the technical field of artificial intelligence, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a target data sample set construction and screening method, device, equipment and medium, and the method comprises the steps: obtaining a target type description, carrying out semantic extension to generate an extension description, generating a reference image sample based on an image generation model, real data units are screened through feature extraction and similarity comparison, and a target data sample set is constructed in combination with knowledge base verification. According to the method, by introducing semantic extension, reference image generation, cross-domain feature comparison and knowledge base consistency verification, real samples highly fitting target type semantics are automatically screened from massive original data, so that the manual annotation dependence is reduced, the illegal sample construction efficiency is improved, and the manual annotation time is shortened. And the training quality and the expansion capability of a subsequent detection model are enhanced.

The invention discloses a semi-supervised text data multi-label classification method, system and device and a storage medium. According to the method, a double-branch model of a shared feature processing network is constructed, a pseudo label generator is utilized to automatically generate a pseudo label for an unlabeled sample on the basis of limited labeled data, and the pseudo label generator and a classifier are jointly trained to realize collaborative learning of labeled data and unlabeled data; by introducing an adaptive threshold mechanism and an improved loss function design, the recognition precision of minority class labels is effectively improved. Compared with a traditional full supervision model, the method has the advantages that the dependence on large-scale manual annotation is reduced, the data preparation cost is remarkably reduced, and the application performance of the classification model in multi-label scenes such as medical text analysis, public opinion monitoring and personalized recommendation is improved. The system, the device and the storage medium provided by the invention can realize the method, and have good expandability and engineering application value.

The invention relates to the technical field of database development, in particular to an SQL generation method and device based on a large model and an ER graph. The method comprises the following steps: analyzing a user query statement to obtain a database candidate table set; querying table field information of each table related to the user query statement from a database candidate table set; according to the table field information of each table, querying an association relationship among the tables in the database candidate table set from a preset ER map, and determining intermediate table information, the intermediate table being used for associating candidate tables without direct association; and generating an SQL query statement according to the intermediate table information and the database candidate table set. Through combination of candidate set generation and SQL optimization, logic errors of the generative large model are reduced, and execution efficiency is improved. The method does not depend on manual annotation of training data of a query language and an SQL template, the labor cost is greatly reduced, and good mobility is achieved.

The invention discloses a patent evaluation method based on a feature-efficacy matrix and a large language model, which performs feature-efficacy matrix analysis and construction on a patent text set to be processed by using text generalization and downstream task potential of the large language model, and does not need to adopt manual annotation, thereby reducing the cost. Wherein at least one feature-efficacy matrix and a patent summary contained in a patent text to be processed are obtained in the analysis and construction step of the large language model. According to the method, a vector and efficacy interval mixed retrieval mode is constructed based on the feature-efficacy matrix, and complex query requirements of patents are accurately matched. Meanwhile, a multi-dimensional patent quality evaluation mode based on a feature-efficacy matrix is a comprehensive evaluation mode from the dimensions of novelty, efficacy, popularity and the like and is applied to patent retrieval, and the analysis effect of related patents is further improved.

This invention relates to a system and method for non-invasive detection of gravesites and archaeological features using multimodal remote sensing and machine learning. Remotely sensed datasets, including RGB, multispectral, hyperspectral, LiDAR, and thermal imagery, are orthorectified, mosaicked, and subdivided into tiled image segments. Features are labeled through manual annotation of visible markers and environmental signatures and expanded via iterative augmentation. A supervised pipeline trains computer vision models, such as YOLO-based detectors, in parallel with tabular models derived from spectral indices (NDVI, NDRE), LiDAR elevation derivatives, and thermal anomalies. Inference outputs are cross-validated against thresholded evidence layers to reject false positives and upgraded when spectral, spatial, and thermal evidence align. Validated detections are exported as GIS-compatible layers with confidence scores and metadata. The system provides a scalable, replicable tool supporting archaeologists, Indigenous communities, and planners in cemetery investigations, cultural resource management, and humanitarian searches for unmarked or clandestine graves.

The invention discloses an encrypted traffic adaptive update classification method and system oriented to an open network environment. The method comprises the following steps: firstly, extracting endogenous semantic features and exogenous environment features based on a causal decoupling mechanism, and stripping an environment confusion factor through anti-fact disturbance and invariance constraint to obtain invariant semantic representation; constructing a macroscopic drift state vector representing a network situation, and inputting a trained meta-learning super-network intelligent decision adaptive control hyper-parameter; performing cross-modal element calibration by utilizing large language model thinking chain reasoning, and calculating the subspace direction consistency of an instantaneous gradient vector of a candidate sample and a category optimization trajectory prototype so as to screen credible samples; and in combination with the capacity-limited playback queue, gradient orthogonal projection constraints are introduced to update low-rank adaptation layer parameters. According to the method, the concept drift problem is solved through causal decoupling and meta-learning decision, forgetting prevention is achieved through orthogonal projection updating, and the online adaptability and robustness of the model in the open environment can be improved without additional manual annotation.

The invention discloses a method for automatically constructing a pathological image data set and training cell nucleus detection and classification based on a space transcriptome technology, and belongs to the field of image processing and artificial intelligence auxiliary pathological diagnosis. According to the method, a spatial transcriptome public data set is obtained, and a data set containing image blocks, weak supervision / semi-supervision labels and cell nucleus boundary information is automatically constructed through preprocessing, deconvolution cell type annotation and cell nucleus instance segmentation, so that the dependence on manual annotation is reduced. Furthermore, a detection and classification model is designed, a multi-scale deformable attention encoder and a decoupled detection and classification decoder are adopted, a limited deformable cross attention mechanism is introduced into the classification decoder, KL divergence classification loss is combined, and instance-level cell nucleus categories are learned from region-level proportion labels. According to the method, end-to-end automation is realized, the cell nucleus detection and classification precision and efficiency are improved, and a high-quality pre-training model basis is provided for downstream pathological analysis.

The invention discloses a semi-supervised image annotation method based on a limited label data set, and the method comprises the steps: taking a FixMatch frame as a basis, and integrating a learnable batch normalization module, a dual-scale parallel convolution module, a content and style separation dual-branch module and a dynamic residual gating module in a ResNet backbone network, the stability of feature extraction and the adaptive capacity to enhanced disturbance are improved. For a label-free sample, a multi-level pseudo-label fusion mechanism is provided, prediction distribution of weak, medium and strong enhanced views is synthesized, and high-confidence pseudo-labels are generated through confidence weighted fusion of multi-level enhanced views and comparison and screening with a category threshold. On the basis, a joint loss function composed of label supervision loss and pseudo label consistency loss is constructed, and a plurality of key control parameters in FixMatch + + are adjusted and optimized in a pre-experiment and grid search combined mode to obtain a group of optimal parameters of the model. Finally, a user inputs a label-free image into the trained FixMatch + + model, and the model can automatically generate a high-confidence pseudo label, so that the number of labeled samples in a limited labeled image set is increased, and the classification precision is improved. By implementing the method, the manual annotation cost can be reduced, and efficient and reliable support is provided for image analysis and recognition tasks.

The invention discloses a scene intelligent extraction method based on L4-level automatic driving minibus road test data, and the method comprises the steps: collecting environment data and vehicle dynamic data, and carrying out the preprocessing, and obtaining time series data; constructing a DSFEM network, inputting time sequence data for training, and constructing a self-supervised loss function to update parameters of the DSFEM network to obtain time sequence features; clustering the time sequence features through a K-means clustering algorithm to obtain clustered scene features; constructing a manual annotation scene library, and calculating the similarity between the scene features to be stored and the behavior features of the existing scenes in the manual annotation scene library; an existing scene with the highest similarity is matched for the scene features, and manual annotation information corresponding to the existing scene is mapped into the scene features, so that scene automatic extraction and manual annotation scene library updating are realized; according to the method, the efficiency is remarkably improved in the aspect of scene generation, the reasonability and the coverage range of a scene library are ensured, and reliable support is provided for testing and evaluation of an automatic driving minibus.

The invention discloses a night no-reference image quality evaluation method based on controllable distortion synthesis, and belongs to the technical field of image processing. The method comprises the following steps: firstly, generating a high-quality night image by using a text generation image model, and introducing various controllable degradation through a preset distortion model to form a distorted image sequence with a pseudo label so as to construct a training set; on this basis, a double-branch hybrid network in which a convolution branch and a Transform branch are parallel is designed, and an adaptive cross-branch feature fusion module is introduced in each stage to dynamically balance global and local features. A meta-learning optimization strategy is adopted to enable the model to rapidly adapt to different distortion tasks, and finally, fine tuning is performed on a small number of subjective scoring samples to obtain an evaluation model. According to the invention, the training data is generated without manual annotation, and the prediction precision and generalization ability of the model in a complex night scene are significantly improved.

The invention relates to the technical field of multi-mode machine learning and music information retrieval, in particular to a video score aesthetics evaluation method based on a cross attention mechanism and weak supervised learning. The invention discloses a video aesthetics evaluation method based on a cross attention mechanism and weak supervised learning, and the method achieves the automatic evaluation of the aesthetics quality of video background music through the construction of a multi-dimensional music aesthetics evaluation network MEMA (Multi-dimensional Evaluation of Musical Athletics). According to the method, a two-stage training strategy is adopted; in a pre-training stage, a model learns a deep corresponding relation between music and a scene through a cross attention mechanism, and manual labeling is not needed; in the fine tuning stage, weak supervised learning is used, a big language model is combined, and a feature projector and scoring ability are trained based on weak tags generated by user comments. The evaluation system comprehensively evaluates the score quality from three dimensions of narrative emotion consistency, technology fusion degree and theme recognition and originality. According to the method, the problems of high subjectivity, high marking cost and insufficient cross-modal understanding of traditional music evaluation are solved, and technical support is provided for applications such as automatic evaluation of video incidental music, incidental music creation assistance and the like.

The invention discloses a prompt word generation method and system based on user habits and documents, and relates to the technical field of natural language processing. According to the method, the user preference model is constructed, the initial cue word set is generated, the real-time document content is matched to screen the target cue word, the cue word is dynamically optimized, the preference model is updated through user feedback, and the cue word recommendation accuracy is effectively improved through collaborative analysis by combining the user behavior data and the document semantic features; the dynamic adjustment mechanism ensures that the system can quickly adapt to new field documents or user behavior changes, the manual annotation cost is reduced through feedback driven model updating, and finally generated cue words not only reserve long-term preference characteristics of users, but also can respond to real-time operation requirements in time, and show higher adaptability in personalized recommendation scenes.

The invention discloses a variable impedance constant force control method for a complex environment based on reinforcement learning, which integrates reinforcement learning adaptive compensation, double-loop force position control and a parameter smooth transition mechanism, and does not need to depend on a large amount of manual annotation data to carry out offline calibration. The damping coefficient in the impedance parameters can be dynamically adjusted in the full life cycle of the robot, and continuous optimization and self-adaptive updating are achieved. Compared with a fixed impedance method, the environmental adaptability is remarkably improved; compared with a traditional self-adaptive impedance method, the self-adaptive impedance method has the advantages that higher convergence speed and higher steady-state precision are achieved in a nonlinear and strong disturbance environment by utilizing the autonomous exploration capability of reinforcement learning, so that the robot can quickly converge to expected contact force in a curved surface, an inclined surface, an uneven terrain and a complex environment; the method has the advantages of being high in response speed, small in overshoot, high in steady-state precision and high in environmental adaptability.

The invention discloses a high-precision remote sensing image cultivated land boundary extraction method, which comprises the following steps of: acquiring high-resolution remote sensing image data, preprocessing the data, and obtaining cultivated land parcel vector data through manual annotation; constructing a multi-task depth semantic segmentation model based on a coding and decoding structure; constructing a mixed loss function; acquiring a high-resolution remote sensing image of a cultivated land area to be detected and performing data preprocessing; inputting a high-resolution remote sensing image of a cultivated land area to be detected into the trained high-resolution remote sensing image cultivated land extraction model based on multi-task learning; and converting the cultivated land semantic segmentation result into cultivated land parcel vector data containing latitude and longitude coordinates. According to the method, through staged Transform coding based on overlapping patches and efficient sequence reduction, efficient multi-scale coding is realized, the problem of insufficient remote dependence capture is solved, and the computing power bottleneck of high-resolution self-attention is relieved.

The invention provides a time series data enhancement method and system for point cloud polar voxel mask modeling, and belongs to the field of deep learning, and the method comprises the steps: S1, employing positioning information to assist point cloud space scanning and sampling, and constructing a basic unlabeled data set; s2, spatial data information complementary increase is carried out through differentiated continuous inter-frame point clouds; s3, improving the unit sampling balance degree through polar coordinate voxelization sampling; and S4, carrying out generative pre-training through the voxel mask modeling network, and improving the generalization perception capability of the neural network on the basic point cloud voxels on the premise of avoiding introduction of complex point cloud manual annotation. According to the method, under the condition that complex manual annotation is not introduced, the generalization perception capability driven by limited sample point cloud data is improved.

The invention discloses an unsupervised deep learning rainfall dry and wet period classification and monitoring method, which adopts dry period CML signal training, learns normal signal features through an auto-encoder, a variational auto-encoder and a fusion model thereof, and realizes rainfall abnormal attenuation detection by taking reconstruction errors and potential distribution deviation as discrimination indexes. According to the method, extra equipment and manual marking are not needed, rainfall events can be automatically identified, cross-link and cross-region generalization ability is achieved, robustness to non-rainfall interference such as a wet antenna effect is high, large-scale real-time monitoring can be achieved in a smart city and a 6G integrated sensing scene, and the method is widely applied to the fields of weather, hydrology and disaster early warning.

The invention discloses a human preference alignment method based on a self-improvement large visual language model. The method comprises the following steps: automatically constructing a preference data set; acquiring an image-question pair from the visual question and answer data set; applying various visual enhancements to the image, and driving a reference model to generate a group of candidate answers in combination with a question; the group of candidate answers are summarized and extracted into a more comprehensive'win 'text, and meanwhile, the text answers of the reference model to the original picture and the question are taken as'fall-fail' texts, so that preference data in an'image-inquiry-'win 'text-'fall-fail' text 'format are formed; secondly, based on the preference data set constructed in the first step, a direct preference optimization algorithm is applied to conduct alignment fine adjustment on a target visual language model, in the fine adjustment process, parameters of a visual encoder are kept frozen, and a low-rank adaptation layer (LoRA) is only introduced into an extended mode alignment module and a language decoder for training; carrying out iterative self-improvement on the model; after fine tuning is completed, taking the optimized model as a new reference model, and repeating the data construction process in the step 1 to generate preference data with higher quality for the next round of optimization; and the circulation is repeated, so that the continuous self-improvement of the model alignment capability is realized. According to the method, self-supervised preference alignment without manual annotation is realized, and a more comprehensive and high-quality'win 'text is generated.

The invention discloses a knowledge base and knowledge graph dynamic access control method and system based on AI classification and grading, and relates to the technical field of data security and access control. Automatically classifying and grading text segments or knowledge graph entities by using an AI model through a data preprocessing module, generating structured tags, and binding the structured tags with original data; after the data are associated and stored to the database through the structured storage module, the authority management module maintains the accessible classification and grade range of the user; and the query filtering module calls the user permission and dynamically filters the retrieval result when the user retrieves, and only returns the authorized content. According to the method, automatic and fine-grained dynamic access control based on data content semantics is realized, the problems of dependence on manual labeling, coarse access control granularity and insufficient flexibility in the prior art are solved, data security and retrieval efficiency are considered, and the method is adaptive to multiple coincidence rules and standards.

The invention discloses a method, a system and equipment for realizing data security classification automation based on a large model, and a medium, mainly relates to the technical field of classification automation, and is used for solving the problems that an existing scheme depends on a rule engine or a simple machine learning model, the performance is poor when unstructured data is processed, and the classification efficiency is poor. The problems that data transmission is delayed and data processing speed is low due to the fact that semantics of data cannot be accurately captured and data needs to be transmitted to a central server to be processed in the prior art are solved. Comprising the steps of inputting to-be-processed data into a trained large language model, and obtaining an output prediction classification result and confidence; obtaining all prediction classification results and confidence in a preset time period, and verifying whether the prediction classification results meet a preset verification mechanism; when a preset verification mechanism is met, determining the predicted classification result as a final classification result; otherwise, the corresponding collected data are uploaded to the manual annotation terminal, and a returned final classification result is obtained.

The present disclosure provides an object positioning method, device, equipment, medium and program product, and relates to the technical field of image processing. The method converts scene data reflecting the geometric information of the surface of an object into a set of local geometric primitives rich in spatial position, geometric type and characteristic parameters, calculates the geometric affinity between the primitives to represent the possibility that they belong to the same object, and then uses a clustering algorithm to realize instance segmentation, thereby solving the technical problem of low precision and poor robustness of traditional object positioning methods caused by occlusion, background interference and a large number of object types in a complex environment, realizing automatic object instance segmentation and positioning based on geometric clues without a large amount of manual annotation, and achieving the effects of improving the recognition accuracy of objects with complex geometric structures, enhancing the generalization ability of the method in real scenes, and significantly reducing the dependence on pre-training data and manual annotation.

The invention provides an engineering data monitoring and analysis method for bridge damage detection. Relates to the technical field of bridge damage detection. The method comprises the following steps: firstly, setting a data set, collecting engineering data of a bridge in normal and damaged states, and carrying out manual labeling; then, cleaning, standardization and feature extraction are carried out on the original data through feature engineering; and then, constructing a hybrid model of a deep neural network and an ARIMA model, which is used for identifying the damage state of the bridge. After model training, damage state identification can be carried out on newly collected bridge engineering data, and a threshold value is set to judge whether the data is abnormal or not, namely whether bridge damage exists or not. According to the method, the automatic feature learning capability of the deep neural network and the time sequence analysis capability of the ARIMA model are combined, and the bridge damage detection efficiency and recognition precision are improved.

The invention discloses a watershed flood type identification method considering a topological relation, and the method comprises the steps: abstracting each hydrometric station in a watershed into graph nodes, and abstracting the upstream-to-downstream communication relation between the hydrometric stations into directed edges, thereby constructing a watershed directed topological graph; for each flood event, standardization processing is carried out on the multi-station features according to nodes and features, and flood identification features are added to each node. Then, a graph auto-encoder based on a graph attention mechanism is adopted to carry out unsupervised reconstruction learning on the graph structure sample so as to extract a graph-level embedded vector; and finally, performing K-Means clustering on the embedded vector, adaptively determining an optimal cluster number according to a contour coefficient, and finally outputting a type label of the flood sample. According to the method, drainage basin flood type supervision information can be generated without manual annotation, and the extracted feature representation can also be used for model training of downstream tasks such as subsequent flood type real-time prediction.

The invention relates to a model prediction result adjusting method and device, electronic equipment and a readable storage medium. The method comprises the steps that query content and context content input into a prediction model are acquired; determining a deviation value of each first lexical element according to the query content, the context content and the initial output content; under the condition that the deviation value of the first lexical unit is greater than a preset deviation threshold value, calculating an updated attention value of the first lexical unit according to the attention value of the first lexical unit and a preset enhancement coefficient; and according to the updated attention value of the first lexical element and the context content, updating the initial output content to obtain target output content. The problems that a common method for eliminating context illusion needs a large amount of computing power and manual annotation, and the development period is long are solved.

The invention belongs to the technical field of data processing, and discloses a task processing method and device, equipment and a storage medium. According to the method, the initial intention is obtained by performing intention recognition on the initial task input by the user, then the target intention is obtained by performing intention correction on the initial intention according to the historical session information corresponding to the user, the processing module corresponding to the initial task is determined according to the target intention, and the initial task is processed through the processing module. According to the method, intention recognition is firstly performed on the initial task input by the user, then intention correction is performed on the initial intention according to the historical session information, dynamic correction can be performed on the initial intention in combination with the historical session of the user, context continuity in multiple rounds of dialogues is ensured, intention misjudgment caused by dialogue fragmentation is avoided, and the user experience is improved. And the initial task is processed through the processing module corresponding to the target intention, so that the task processing accuracy is improved, a large amount of manual annotation data is not needed, and the task processing cost is reduced.

According to the multi-modal pedestrian re-identification method based on fine-grained attribute enhancement and noise perception provided by the invention, the fine-grained attribute is enhanced for re-identification of general guidance personnel based on a multi-modal large language model, automatic enhancement of the fine-grained attribute is realized, the manual annotation cost is reduced, and high-quality text data covering multi-dimensional details is generated; and meanwhile, the generated matching score also provides a reliable quantitative basis for subsequent noise filtering, and also provides rich semantic supervision signals for cross-modal alignment, so that the adaptability of the model to cross-camera visual angle difference is integrally improved. In addition, through comparative learning between images and texts and feature alignment research between different modal data, a fine-grained feature alignment framework of noise perception is constructed, a text credibility scoring mechanism is introduced, low-credibility samples are filtered, interference of noise on model training is suppressed, and cross-modal feature alignment precision and model robustness are improved.

This invention relates to a global feature tracking method for time-varying data based on unsupervised segmentation. It constructs a volume data segmentation network to segment the input volume data and achieve global tracking. This invention introduces deep learning to segment volume data, separating feature domains from the background. This enables automatic and accurate tracking of specific features in complex data without any manual annotation, reducing tracking complexity and improving tracking accuracy. It introduces a global feature tracking method; by selecting a target feature, it can track the trajectories of features similar to the target feature across all time steps. Users can select any feature at any time step to track features of interest in time-varying data. It can track extracted features from a global perspective, avoiding tracking errors and defects caused by local tracking methods. Furthermore, it adds the tracking of spatiotemporally similar features, simultaneously tracking the complete paths of spatially similar features of the target feature, improving feature tracking accuracy.

This invention provides a vision-based self-supervised off-road terrain accessibility estimation method. This method utilizes measurement data from an onboard inertial measurement unit to generate self-supervised access cost labels, enabling the predicted terrain cost to reflect the dynamic response characteristics of the vehicle during driving. Furthermore, these labels do not require manual annotation, saving the cost of training data labeling. Images containing rich terrain features are used as input perceptual information for the prediction network, and a contrastive learning method is used to pre-train the backbone network that extracts multi-scale features from the images, improving the reliability and accuracy of off-road terrain accessibility estimation.