19 results about "Classification structure" patented technology

This invention discloses a tracking method based on multi-scale spatial constraints to combat occlusion. It extracts HOG, CN, and grayscale features from candidate target regions and uses PCA to reduce the dimensionality of HOG and CN features, accelerating computation. A channel weight fusion method is employed, training filters separately for each feature layer. Adaptive fusion weights enhance the response of effective feature layers, addressing the problem of insufficient multi-feature response fusion. A stepped spatial constraint method is proposed to optimize the color space constraint model, preventing model errors from obscuring target information and limiting the effectiveness of the spatial domain constraint model. Adaptive learning rate and diffusion search methods reduce irrelevant information learned by the filters and improve tracking accuracy when the target is occluded. A tritree scale acceleration method is proposed, introducing scale filters and transforming the parallel structure of the scale filters into a tritree classification structure.

The invention provides a data classification management device for educational management. The data classification management device structurally comprises a power-on control end, an electronic storage main body, a connection frame, a heat dissipation end and a paper storage structure, after a paper storage structure is improved, paper materials are classified and stored in order through the storage grooves of the storage box and the classification structure, then mutual communication of the ventilation end and the heat dissipation end is ensured by combining position positioning of the insertion blocks and the power-on rods, and then the ventilator is started to lead out humid and turbid air in the storage grooves; furthermore, hot air blown out from the heat dissipation end can enable the interior of the storage groove to be in a continuous dry state, so that the condition that fonts are fuzzy due to the fact that paper files are affected with damp due to continuous contact with outside air can be effectively avoided; and meanwhile, the classification structure can improve the classification searching efficiency of the paper files through the ventilation clamping sleeve and the label groove of the supporting block, the empty groove position receives hot air of the heat dissipation end, and classification protection of the paper files is improved.

The application discloses a kind of defect grading method, device, equipment and storage medium.The method comprises: obtaining defect image to be graded;The defect image to be graded is input to the grading model pre-trained;According to the output result of grading model, the defect level of defect image to be graded is determined;Wherein, the grading model is obtained by training the grading network pre-constructed based on defect grading dataset;The grading network includes feature extraction structure, regression structure and classification structure;Defect grading dataset includes at least one group of defect grading data, and each group of defect grading data includes defect image sample and defect image sample matched reference level;The number of reference level is less than the number of defect level.This technical solution solves the problem of low defect grading refinement degree, can realize fine-grained defect grading while effectively improving the accuracy of defect grading to meet the diversified workpiece needs of users.

The embodiment of the invention provides a classification model training method based on hierarchical knowledge migration. The method comprises the steps of obtaining a universal baseline model with a semantic understanding capability; according to the hierarchical classification structure of the target domain, an industry domain model containing a plurality of sub-classification output heads is constructed, and initialization is carried out by using a general baseline model parameter; constructing a hierarchical constraint loss function, and constraining the parent class prediction probability to be greater than or equal to the child class prediction probability; obtaining domain classification training samples, and respectively inputting the domain classification training samples into the general baseline model and the industry domain model to obtain first and second prediction category probability distributions; calculating domain cross entropy loss based on the second prediction distribution and the real label, calculating knowledge distillation loss based on the difference between the two distributions, and constructing a target loss function in combination with hierarchical constraint loss; and taking minimization of the target loss function as a target training industry domain model. According to the method, the model training efficiency is effectively improved, efficient migration of domain knowledge is realized, and the hierarchical classification accuracy is remarkably improved.

The application relates to the technical field of natural language processing, and discloses a history material text automatic classification method based on natural language processing. The method comprises the following steps: collecting original history material text data and preprocessing; constructing a text feature graph containing multiple text entity nodes, semantic correlation relationships between entities and feature information of each node; identifying core semantic clusters, discrete semantic fragments and positioning key entity nodes in the graph through a text analysis algorithm; generating a preliminary classification strategy set containing a semantic remapping scheme, a classification structure adjustment scheme and a weight adjustment scheme according to the core semantic clusters and the key entity nodes; receiving a user correction instruction, determining a classification correction type and adaptively adjusting the preliminary classification strategy set; and executing the adjusted strategy set, and updating the text feature graph according to the classification result. The method improves the classification accuracy and flexibility of history materials, and helps efficient arrangement and utilization of history material resources.

The invention discloses a fireproof control chart auxiliary drawing method based on an intelligent model, and the method comprises the steps: manually obtaining all drawing files related to a fireproof control chart, converting the drawing files into a picture format in batches through a Python bitmap conversion method, and keeping the classification structure of the drawing files. The drawing file in the picture format is converted into text data, and prompt words are written based on the text data. And creating a calling variable of the model, and transmitting the calling variable and the cue word to the graphic understanding model to obtain an analysis result in a fixed format. And inputting the analysis result in the fixed format and the cue word into an inference model to extract json key data information, and converting the json key data information into a json key data table by utilizing Python. And establishing connection with the drawing software by utilizing Python, and drawing in the drawing software in sequence according to the json key data table. The working efficiency of drawing is improved, working hours are saved, and the labor cost of designers is reduced.

The invention discloses a digital semantic understanding and conversion system and method based on multi-modal feature fusion, the contribution degrees of three types of features including semantics, mathematical and context in different digital type recognition are displayed through a multi-modal feature importance thermodynamic diagram, a macroscopic-microscopic double-layer classification structure is adopted, error propagation is effectively blocked, and the recognition efficiency is improved. Through three-way cooperation of a rule engine, a statistical model and a semantic engine, high-robustness decision making is realized, and an intelligent conversion routing and grammar adaptation mechanism is developed for a digital semantic conversion link. The method can effectively solve the technical bottlenecks of insufficient accuracy, poor expandability, strong training data dependence and the like of a traditional method in complex number type recognition, and is suitable for multiple application scenes such as TTS voice generation, natural language generation, financial document processing, education science and technology and the like; accurate conversion from digital character strings to semantic expressions conforming to human language habits can be achieved.

This invention discloses a method and apparatus for intelligent identification and classification of industrial assets based on a large language model. The method includes: actively probing the internet address space and collecting web page data; generating page clusters using a two-stage clustering approach that combines structure and semantics; performing semantic summarization and multi-round filtering using a large language model to refine data related to industrial control systems; constructing multi-dimensional feature summaries and generating atomic tags; and generating a hierarchical classification chain based on a dual-encoder retrieval enhancement generative model. This invention achieves automatic discovery and refined classification of industrial control system assets through multi-dimensional semantic fusion and retrieval enhancement reasoning, solving the problems of traditional methods such as reliance on strong rules, low recognition accuracy, and lack of classification structure.

The utility model provides a sample sorting device with a multi-classification structure, which comprises a sorting device body, a conveying assembly and sorting assemblies, the left upper side and the right upper side of the interior of the sorting device body are respectively provided with one sorting assembly which is used for classifying and sorting collecting tubes on the conveying assembly based on mechanical vision. Compared with the prior art, the collecting tube sorting device has the advantages that the sorting windows are arranged on the front side of the shell, the storage boxes with the number corresponding to that of the sorting windows are installed, a large amount of collecting tube sorting work can be processed when necessary, various classification selections are provided, various collecting tubes can be classified and sorted in a more detailed mode, and the sorting efficiency is improved. The conveying assembly and the sorting assembly are arranged, and the visual assembly and the double sorting assemblies are matched, so that a large number of collecting pipes can be processed at the same time, the equipment has the capacity of processing a large number of collecting pipes, and the requirement for sorting the collecting pipes of various classifications can be met.

The present application relates to the technical field of deep fake detection, in particular to a deep fake detection method based on local-global self-supervised contrastive learning; the method comprises the following steps: firstly, collecting unlabelled face images and labelled real-fake data and uniformly preprocessing, including face detection, extension cropping and key point positioning; secondly, dividing two global perspectives and seven local perspectives according to the key points, and constructing multi-view contrastive learning samples; further, designing a self-supervised pre-training process based on a teacher-student framework, learning robust face representation unsupervisedly by comparing the consistency of local and global features and reconstructing mask image blocks; finally, supervisedly fine-tuning the pre-trained model, and optimizing the classification structure by using neural network architecture search. The present application realizes high-precision and strong-generalization deep fake detection.

The application discloses an intrusion detection method and system based on integrated under-sampling and adaptive parameter optimization, randomly selects a hyperparameter combination of a base classifier in a hyperparameter space, obtains a globally optimal parameter combination through quantitative scoring, trains a temporary integrated model composed of temporary base classifiers, and selects an optimal weight combination through performance evaluation; in t rounds of iterations, a current integrated model is constructed by using the base classifiers generated in the previous round of iteration, majority class samples are predicted and divided into bins, the majority class samples are under-sampled according to the misclassification rate, and the under-sampled majority class samples are combined with minority class samples to form a balanced training set of the current round; based on the globally optimal parameter combination, the optimal weight combination and the balanced training set, the base classifiers of the current round are trained, the comprehensive weights are calculated, and the base classifiers are added into the final integrated model; the application can retain boundary majority class sample information, adaptively optimize the base classifier structure, and significantly improve the classification performance and detection rate of a network intrusion detection system.

The application discloses a kind of multi-omics data hierarchical classification structure learning systems based on machine learning, comprising: data import module, for loading multi-omics data, and the data is preprocessed;Hierarchical structure learning module, using the adaptive machine learning method driven by data carries out supervised learning, constructs category similarity matrix, uses bottom-up, top-down unsupervised clustering algorithm to preliminarily construct class label hierarchy, and integrates iterative algorithm into training process, finally obtains optimal class label hierarchy;Hierarchical classification verification module, using optimal class label hierarchy combines multi-omics data to carry out hierarchical classification, and provide result explanation.The application combines supervised, unsupervised machine learning and iterative algorithm, accurately infers the class label hierarchy in multi-omics data, improves the efficient classification and organization of complex multi-omics data, and provides classification result interpretability.

The invention discloses a multi-label data classification method and device based on label noise and hierarchical modeling, and belongs to the technical field of text data processing. The method comprises the following steps: obtaining to-be-classified text data, wherein the to-be-classified text data has a tag semantic hierarchical feature; inputting the to-be-classified text data into the trained classification model to obtain classified data; the trained classification model comprises a feature extraction structure, a classification structure and a three-level label verification structure, the feature extraction structure is used for extracting heterogeneous features of to-be-classified text data in different subspaces, and the classification structure is obtained by training the classification model by using a positive sample-first collaborative noise reduction strategy for noisy label data; the label semantic layering features comprise a first-layer macroscopic label feature, a second-layer specific label feature and a third-layer related label feature, and the three-level label verification structure is used for verifying the first-layer macroscopic label feature, the second-layer specific label feature and the third-layer related label feature.

The present application relates to a kind of model-oriented geometric body and three-dimensional mark information extraction and reorganization system, belong to product data management technical field.The system includes: integrated adaptation module, as front-end plug-in is deployed in three-dimensional CAD software, for receiving user instruction and establishing communication with PLM system;Intelligent extraction module, calling CAD kernel interface directly reads the geometric information of model, assembly structure, attribute and three-dimensional mark, and is intelligently identified and classified by rule set;Structured reorganization and incremental synchronization module, the data extracted are encapsulated into composite object according to PLM data model, and only difference part is incrementally transmitted by comparing version or content hash;Lightweight generation and association module, the geometric data containing mark are converted into lightweight model that Web can render, and the metadata associated with PLM business object is embedded.The present application realizes the accurate, efficient, intelligent synchronization of design data to management system.

This invention discloses an intent recognition method and system based on a generative large language model, belonging to the field of natural language processing. It solves the problems of insufficient generalization ability, high manual annotation cost, and small context window of the traditional BERT model for intent recognition. This invention retains the traditional NLP supervised classification idea, modifies the Qwen3-0.6B large language model, and adds a classification head to construct an intent classification model. The model is constructed through data processing and fine-tuning during training. In application, user input containing single-turn / multi-turn context is collected. The large model extracts semantic features, the classification head judges and outputs the intent confidence score, and outputs the valid intent after a 0.95 threshold validation. If the score is not met, historical dialogue is concatenated for re-identification. This invention achieves decoupling and fusion of deep semantic understanding and traditional classification structure, improving recognition accuracy, reducing annotation cost, and enhancing system controllability and scalability. It is suitable for multi-turn dialogue scenarios such as intelligent customer service.

The invention discloses an intention recognition method and system based on a generative large language model, belongs to the field of natural language processing, and solves the problems that a traditional BERT model is insufficient in intention recognition generalization ability, high in manual annotation cost, small in context window and the like. According to the method, a traditional NLP supervised classification thought is reserved, a Qwen3-0. 6B large language model is transformed, a classification head is additionally arranged to construct an intention classification model, and model construction is completed through data processing and training fine tuning; during application, user input containing single-round / multi-round contexts is collected, semantic features are extracted through the large model, classification head discrimination is performed, intention confidence is output, effective intentions are output after 0.95 threshold verification, and if the intentions do not reach the standard, historical dialogues are spliced for re-recognition. Decoupling fusion of deep semantic understanding and a traditional classification structure is achieved, the recognition accuracy is improved, the labeling cost is reduced, the controllability and expansibility of the system are enhanced, and the method is suitable for intelligent customer service and other multi-round dialogue scenes.

The invention discloses an artificial intelligence visual system based on semantic archiving and an application method thereof, the system breaks through the calculation limitation of a traditional end-to-end generation model, and efficient content generation and recognition are achieved through modular visual fragment recombination. The system comprises a data acquisition unit which acquires multimedia data from various heterogeneous data sources, wherein the multimedia data comprises a static data set, a real-time video stream and 3D rendering data; the semantic segmentation unit adopts edge detection and a clustering algorithm to decompose multimedia data into visual fragments with semantic meaning, and the visual fragments comprise spatial metadata, confidence information and edge compatibility descriptors; the vector quantization unit converts the visual fragments into compact numerical vector representation by extracting a color histogram, LBP texture features and geometric moments; the hierarchical labeling unit adds multi-level semantic metadata and technical metadata for the visual fragments, the semantic metadata is organized according to a hierarchical classification method structure, and the technical metadata defines contact edges and assembly rules to support intelligent recombination; the semantic technology archiving unit adopts a double index mechanism to store and retrieve visual fragments with metadata, and supports hash retrieval based on semantic tags and similarity search based on technical descriptors; the multi-modal retrieval unit uses a semantic-technology score fusion algorithm to sort the candidate segments; the intelligent assembly unit recombines the selected fragments based on predefined topology rules and consistency constraints; the selective refinement unit applies GPU-assisted local optimization processing only at the fragment connection points. According to the system, a new visual processing normal form based on semantic memory and fragment modularization recombination is achieved, the calculation load is remarkably reduced, the large-scale retraining requirement is eliminated, and efficient visual generation and recognition in a real-time application environment are supported.

The invention discloses a self-adaptive service navigation method, device and system and a computer readable storage medium, and belongs to the technical field of artificial intelligence. The self-adaptive service navigation method comprises the following steps: acquiring to-be-processed text data input by a user; performing semantic analysis on the to-be-processed text data by adopting a field classification model based on an attention mechanism, and determining an input type corresponding to the to-be-processed text data: analyzing the to-be-processed text data based on a pre-constructed hierarchical intention classification model under the condition of determining that the input type represents a business requirement type, positioning a target service node in the intention classification structure of the full-service menu; and based on the target service node and the menu mapping rule of the target navigation scene, determining a target service and performing navigation. Therefore, classification accuracy and positioning precision are improved, business menu logics of different scenes can be quickly adapted, the difficulty of cross-mechanism deployment is reduced, flexible popularization and efficient configuration are realized, and scene adaptation flexibility is enhanced.

The utility model discloses a classification structure of a glass bottle mouth defect inspection machine, which belongs to the technical field of inspection machines, and comprises an inspection machine body and a control host arranged on the inspection machine body, a conveying frame is transversely arranged on the inner side of the inspection machine body in a penetrating manner, and the conveying frame is arranged on the inner side of the inspection machine body. The two ends of the conveying frame are rotationally connected with a driving roller and a driven roller correspondingly, and the side wall of the conveying frame is provided with a gear motor with the power output end in transmission connection with the driving roller and a conveying belt in transmission connection between the driving roller and the driven roller. Glass bottles with bottle openings having defects are moved to the edge of the left side of the conveying belt and continuously conveyed by driving the classifying plate to rotate anticlockwise through the servo motor, the two types of glass bottles are separated and classified through the partition plate, and a worker is located at the tail end of the conveying belt 15 to collect the glass bottles having the defects. And recycling is carried out.