38 results about "Linear classifier" patented technology

The invention discloses a handwritten text recognition method based on multi-stage enhancement. The handwritten text recognition method comprises the following steps: acquiring a handwritten text image; constructing a hierarchical dynamic multi-scale CNN backbone network to obtain a visual feature sequence; inputting the visual feature sequence into a time sequence multi-scale module to obtain a local enhanced feature sequence; performing global modeling on the local enhanced feature sequence by using a Transform encoder to obtain a global visual feature sequence; enhancing the global visual features to obtain a time sequence context feature sequence; dynamic weighted fusion is carried out on the global visual features and the time sequence context features through a gating fusion module; and sending the fused features into a linear classifier and a CTC decoder to obtain an identification result. According to the method, intelligent arbitration of global and time sequence features is realized through new technology application of a hierarchical multi-scale CNN trunk, time sequence context enhancement and a gating fusion mechanism, and the recognition accuracy and robustness are remarkably improved.

The application discloses a kind of light-weight human key point detection method and device based on model pruning, comprising the following steps: (1) the original image is convolved and continuously down-sampled, and different scale low-resolution images are obtained;(2) a multi-branch convolutional neural network is constructed, different resolution images are input into branch network respectively, and information fusion layer is set between different branches;(3) a linear classifier is added to each convolution block in the network, and the contribution of each convolution block is calculated using the accuracy of the linear classifier;(4) the contribution of the convolution block is sorted, and the convolution block with low contribution is deleted according to the pruning ratio to obtain a pruned network;(5) the pruned network and the original convolutional neural network are trained together using the knowledge distillation method;(6) the image to be detected is input into the trained pruned network, and the detection result of human key points is obtained.The application can effectively improve the accuracy and reduce the model parameters.

The invention discloses a network flow classification method and device based on multi-modal feature fusion, and the method comprises the steps: obtaining network flow data, extracting an IP address as a node, and constructing a communication graph; time modal features and event modal features are extracted for each node, and standardization processing is carried out on the time modal features and the event modal features; the time modal features and the event modal features are aligned; fusing the two types of aligned features; performing multi-layer neighbor sampling and hierarchical aggregation on the communication graph by using a graph neural network model to obtain structure-enhanced node features containing multi-hop neighbor information, inputting the node features into a linear classifier, outputting a category probability, selecting a category with the maximum probability as a prediction result, and performing iterative training to obtain a classification model; and mapping the classification model to a data plane of the programmable switch to realize online reasoning of the data packet. According to the method, high-precision identification of various network traffic types is realized by fusing the spatial modal, time and event modal characteristics.

The invention discloses a alumen ustum image clear area screening method based on superpixel segmentation and feature classification. The method comprises the following steps: generating a definition image of an original alumen ustum image by using a Laplace operator; enhancing the color contrast of the original alumen ustum image by using the definition image; segmenting the enhanced image into a plurality of small regions by using an SLIC superpixel segmentation method, and storing segmentation boundaries of all the small regions; applying the segmentation boundary to the original alumen ustum image, filling all small areas into a minimum enclosing rectangle, and storing the position and structure information of the minimum enclosing rectangle; using a ResNet feature extraction network to perform feature extraction on the small regions obtained by segmentation; and performing definition classification on all the small areas by using a linear classifier, and returning a classification result to an original image to obtain a clear alumen ustum image. According to the method, automatic identification and screening of clear alumen ustum structures in the image are realized through image small region division and region feature discrimination.

Systems and methods for property guided molecular optimization using artificial intelligence diffusion models. An equivariant continuous denoising diffusion implicit model autoencoder framework (DDIM-AE) can be trained (510) on a conformational dataset to predict raw data from data corrupted by a time-dependent noise to obtain a trained DDIM-AE that ensures controlled generation of threedimensional (3D) molecules. Linear optimization of semantic embeddings of 3D molecules can be performed (520) with a linear classifier to achieve a target property value from desired properties and obtain an optimized embedding. An optimized 3D molecule that includes molecular conformation with the desired properties while preserving interactions with biochemical molecules can be generated (530) from the optimized embedding with the trained DDIM-AE.

The invention relates to a time-space diagram neural network autism classification method based on dynamic function connection and dynamic effective connection feature fusion, which can realize accurate classification of autism by using resting state functional magnetic resonance imaging data. The method comprises the following steps: firstly, constructing a dynamic function connection matrix and a dynamic effective connection matrix to respectively extract brain network diagram characteristics, and under the guidance of a dynamic effective connection network, forming fused brain network space-time connection characteristics by adopting a space-time fusion position Transform based on a cross attention mechanism; introducing a multi-layer perceptron to extract high-order image features in the brain network, embedding the high-order image features as node representation of a population graph, constructing edges of the population graph by using demographic information, realizing fusion of the magnetic resonance image features and the demographic information, and finally learning node embedding through a graph convolutional network to obtain the demographic information of the population graph. And autism classification is realized based on a linear classifier. Experimental results show that the provided method has excellent performance in autism diagnosis tasks, and the accuracy and robustness of diagnosis are remarkably improved.

Methods for obtaining black-box representations of machine learning models are disclosed when information about the models' internal states or parameters is inaccessible. By using the model's outputs instead of its internal states, the black-box representation is model-agnostic and provides a reliable and robust representation of the model through an external lens. The black-box representation is generated using responses from the model to a series of initialization and information-gathering questions, quantifying the model's confidence in the responses it has just returned. The black-box representation is then used as a training dataset for a linear classifier to learn performance metrics about the model.

The application discloses an industrial information physical system cross-layer anomaly detection method based on information physical feature fusion, acquires a data packet sequence of each communication event; for each data packet sequence, extracts coarse-grained features of a sequence level thereof, analyzes all data packets to obtain fine-grained features of a data packet level, and then fuses the coarse-grained features and the fine-grained features to obtain an overall feature vector of the sequence; taking the overall feature vector of the data packet sequence as observation data, projecting the observation data to a high-dimensional feature space by using a dictionary, learning the dictionary and a linear classifier; in online testing, obtaining an overall feature vector of a data packet sequence corresponding to a current communication event, obtaining sparse coding based on the dictionary, and outputting a current state label of the industrial system by using the linear classifier. The application improves real-time performance and accuracy of anomaly detection by effectively fusing information physical heterogeneous data.

The invention provides a method and a system TGrKCMI for predicting interaction between circRNA (Ribonucleic Acid) and miRNA (Micro Ribonucleic Acid). According to the method, a pre-training model is used for extracting sequence features of circRNA and miRNA, and redundant information is reduced through PCA dimension reduction; then, a multi-head attention module with a gating mechanism is introduced to encode sequence features, and the expression ability of key information is enhanced; on the basis, a circRNA-miRNA interaction diagram is constructed, and robust diagram feature learning is realized in combination with diagram attention convolution of random feature masks. And finally, a kernel-based adaptive nonlinear classifier is adopted to carry out modeling on the fusion features, and high-precision and high-robustness interaction prediction is realized.

The application relates to an end-to-end attribute-level sentiment analysis method based on a sentence component perception attention mechanism and belongs to the technical field of natural language processing. The application comprises the following steps: firstly, a pre-training language model is used to train a word vector. Secondly, a sentence component perception attention mechanism C-ATT is used to divide sentence components, words belonging to the same component in a sentence divided by the C-ATT pay attention to each other, and words of different components do not pay attention to each other, so that aspect words pay more attention to opinion words divided into the same component. Finally, a linear classifier is used to predict a joint label; experiments are conducted on four public sentiment analysis data sets, and the experimental results show that the end-to-end attribute-level sentiment analysis method based on the sentence component perception attention mechanism is effective.

This invention discloses an interpretable robotic arm bearing fault diagnosis method based on physical prior knowledge, belonging to the field of rolling bearing fault diagnosis technology. This method integrates the physical prior knowledge of the bearing with a neural network structure. It obtains fault frequency harmonic features with clear physical meaning and key frequency band broadband features through parallel physical feature extraction networks and broadband feature extraction networks, respectively. A multi-head attention mechanism with sparsity and diversity constraints is used to adaptively filter the fused features, highlighting key fault characterization information. Finally, a single-layer linear classifier achieves transparent fault category decision-making. This invention solves the "black box" problem of existing deep learning fault diagnosis models, providing end-to-end interpretability from feature contribution and attention allocation to classification decision-making while ensuring high diagnostic accuracy, significantly improving the model's reliability and engineering practicality.

This invention discloses a method, system, electronic device, and storage medium for processing chest images, relating to the field of image processing. The method includes: acquiring a multi-label real dataset and a single-label synthetic dataset; dynamically sampling and matching synthetic data batches based on the label distribution of the real data batches; inputting the two types of data batches into a dual-branch heterogeneous framework for cross-modal deep fusion to extract real and synthetic fused representations; using a category-specific parallel linear classifier for feature decoupling and outputting the prediction probabilities of each branch; constructing a mask based on multi-label and single-label vectors to calculate label-level collaborative distillation loss, and combining this with cross-entropy supervised loss to perform end-to-end training and optimization of the model; finally, using the trained model to output feature label prediction results for the image to be processed. This invention solves the algorithmic challenges of feature entanglement caused by the coexistence of multiple attributes and the scarcity of rare samples under long-tailed distributions, improving the accuracy of classification models in extracting pure features.

This invention relates to the field of artificial intelligence technology and discloses a speech extraction method, apparatus, device, and medium. The method includes: first, acquiring reference speech of the target speaker and mixed speech of all speakers; preprocessing and encoding the reference speech and mixed speech to generate two discrete token sequences; fusing the two discrete token sequences to form a fused discrete token sequence; using a language model to predict the fused discrete token sequence to generate candidate discrete token sequences for the target speaker; calculating the probability distribution of the candidate token sequences using a linear classifier and selecting the sequence with the highest probability as the target discrete token sequence; and then reconstructing the target discrete token sequence into a speech waveform to obtain the speech of the target speaker. This invention transforms the complex audio generation problem into a classification problem, simplifying model training; and utilizes the sequence modeling capability of a language model to capture long-term dependencies between speech tokens, achieving high-quality speech reconstruction.

Embodiments of the present disclosure provide functionality to tensor network framework designed for quantum kernel methods and demonstration of tensor network effectiveness at scaling this application. Quantum kernels capture the distance between data points in quantum feature space by evaluating the quantum state overlaps associated with each data point. It has been found that expressing data in quantum feature space may produce more separable data that improves the results of linear classifiers. The different kernel elements may be computed independently, and parallel processing may be exploited to significantly reduce computational time, enabling to train on more data. Thus, quantum kernels continue to improve classification metrics with the addition of more training data and more features.

The application discloses a kind of electroencephalogram emotion signal recognition method based on graph regular non-negative matrix decomposition, comprising the following steps: step one, using nearest neighbor method for electroencephalogram emotion signal constructs an adjacent matrix;Step two, establish a graph regular non-negative matrix decomposition model, and the non-negative constraint of matrix after decomposition is carried out;Step three, introduce a projection matrix, further to matrix implement three decomposition;Step four, the model is optimized;Step five, using the new representation PX obtained after the multiplication of trained model parameter P and sample X replaces the representation matrix V of original model, joins corresponding label matrix and trains a classifier in linear SVM classifier;Step six, the class of test sample is predicted by inputting into classifier.The application can effectively identify electroencephalogram emotion signal, compared with other classic electroencephalogram emotion recognition method, the application effectively improves recognition rate.

A system and method are provided for implementing transformer-based computer vision models on resource-constrained devices. An input image is divided into tokens, each corresponding to a patch. A background-aware vision transformer (BAViT) classifies tokens as foreground or background using a lightweight architecture without a class token and with a linear classifier for token-wise prediction. Training utilizes an accumulative cross entropy loss that aggregates token-level losses to improve accuracy. Tokens classified as background are pruned, thereby reducing computational complexity, runtime memory, and inference latency. Foreground tokens are processed in a downstream transformer-based object detection model, such as YOLOS, to generate detection outputs. The BAViT module operates as a pre-processing stage, facilitating integration with detection models without retraining. Configurations include BAViT-small with two transformer layers suitable for edge devices, supporting applications such as security and inventory tracking.

The invention discloses a trajectory collision risk judgment method, system and device based on multi-dimensional state features and linear separability, and relates to the technical field of intelligent traffic and machine learning. The method provided by the invention comprises the following steps: acquiring a multi-dimensional state vector set corresponding to each of two moving entity tracks, wherein the multi-dimensional state vectors are used for representing the space occupation state of a moving entity at each timestamp; respectively marking the state vector sets of the two tracks as a first category and a second category, and constructing a feature data set; a linear classifier is adopted to process the data set, and a judgment index representing the linear divisible degree of the two categories is generated; and finally, judging whether a collision risk exists or not based on the judgment index, and if the judgment index shows that the collision risk is completely linearly separable, judging that the collision risk does not exist. According to the method, a complex time sequence geometry problem is converted into an efficient high-dimensional data classification problem, the calculation complexity is remarkably reduced, and the efficiency and the real-time performance of large-scale trajectory data analysis are greatly improved.

This invention provides a few-shot action recognition network, method, device, and storage medium based on meta-learning and time alignment. It includes a meta-training unit and a meta-testing unit. The meta-training unit trains an embedding model with good generalization performance to extract temporal and spatial information from video frames. After training, all network headers, including the classifier, are removed. η The embedded model / feature extractor, whose parameters are frozen or fine-tuned as a meta-test unit, is then evaluated by a linear classifier. This invention applies the novel meta-learning algorithm NIL to few-sample action recognition, achieving accurate and efficient classification of human actions with limited sample data.

The application discloses an anaerobic fermentation instability early warning method based on near-infrared transmission spectrum, which collects near-infrared transmission spectrum data of biogas slurry in an anaerobic fermentation impact test and corresponding biogas slurry samples online, determines the spectrum data corresponding anaerobic fermentation state according to the biogas slurry index determination results and the fermentation state threshold of the propionic acid / acetic acid ratio and the volatile fatty acid / total alkalinity ratio, performs wavelength variable primary selection on the preprocessed calibration set spectrum data by using linear classifier integrated wavelength selection fusion mutual information nonlinear feature supplement, simultaneously optimizes the characteristic wavelength and the extreme gradient boosting modeling hyperparameter by using a genetic annealing particle swarm algorithm, and constructs an anaerobic fermentation state intelligent diagnosis model, so that the anaerobic fermentation instability early warning can be completed in combination with host computer software and a 4G SMS cat. The method has unmanned online monitoring function, can realize more than 3-minute sampling interval full-automatic anaerobic fermentation state intelligent diagnosis and instability early warning, and further ensures efficient and stable operation of a biogas project.

The present application relates to the technical field of artificial intelligence, and proposes a two-stage fine-tuning and decoupling reasoning method and device of a visual language model, which comprises: in the first stage, panoramic view and subject view are obtained through center cropping, and a learnable prompt vector is initialized for each view, the encoder is frozen, and the prompt vector is optimized to decouple view-specific semantics; in the second stage, the prompt vector is frozen, the text embedding stack is stacked as a trainable weight matrix, the linear classifier is optimized to enhance the discriminability of the base class, and the optimal fusion weight is searched on the validation set; the decoupling reasoning of the third stage comprises fusing the logits of the two views using the optimal weight fusion for the base class, adopting Dempster-Shafer evidence theory fusion for new classes, and introducing an uncertainty measure to realize robust prediction. The present application effectively balances the recognition performance of the base class and the generalization ability of the new class, and exhibits superior accuracy and robustness in a multi-modal classification task.

The invention provides an SLCC converter fault detection method based on multi-module cooperation, and belongs to the technical field of power systems, and the method comprises the steps: collecting the electrical parameter data of each module of an SLCC, calculating a coupling parameter, and constructing an incidence matrix; dynamically updating the incidence matrix to obtain an updated incidence matrix; an SLCC topological graph is constructed; on the basis of the incidence relation of the topological graph, node features are updated through an iterative formula, and final target features are obtained; inputting the final target features into a trained linear classifier, and judging the fault type at the node i and at the (t + 1) th moment by the linear classifier; and continuing to obtain the updated incidence matrix along with the time, and judging the fault type at the node i at the next time through the linear classifier. According to the invention, accurate tracing of fault causes of multi-module cooperation is realized, and the problem that only a single fault module can be positioned and cross-module cascading fault causes cannot be identified in the prior art is solved.

This application discloses a method and device for diagnosing crack faults in water pump turbines based on neural networks, belonging to the field of water turbine fault diagnosis technology. The method includes: acquiring real-time vibration data of the water pump turbine to be monitored; inputting the real-time vibration data into a trained self-supervised crack diagnosis model to obtain crack diagnosis results; the self-supervised crack diagnosis model includes an encoder, a decoder, and a linear classifier; determining whether the crack diagnosis results are normal; if the crack diagnosis results are normal, inputting the corresponding real-time vibration data as input parameters into a PCA-PSO-BP neural network to obtain the crack initiation location and crack propagation direction; if the crack diagnosis results are abnormal, outputting a fault alarm. This method reduces the difficulty of early crack feature extraction, eliminates the need for a large number of labeled fault samples, reduces the computational cost of full-field stress analysis, and achieves integrated monitoring from intelligent diagnosis to crack evolution prediction.

The application discloses a cross-domain intelligent wireless sensing method based on federated learning and block chain, relates to the technical field of wireless sensing, solves the problems of efficient cooperation and privacy protection of cross-domain heterogeneous data, insufficient generalization ability of multi-target sensing model under a dynamic environment, and contradiction between real-time sensing and calculation efficiency under resource constraints, and the application constructs a FL-BLC cooperative security framework based on lightweight federated learning and block chain to perform distributed training on a local client model; on the basis of the framework, a DB-SE-Yolov8 sensing algorithm model is constructed to extract and reserve global and local fine-grained information by adopting a parallel double-branch network structure, to dynamically weight and fuse multi-scale feature information through a gating mechanism, to perform regression statistics on the extracted feature information by using a full-connection linear classifier, to perform probability mapping by using a Softmax activation function, and to output a sensing result; while improving the precision of cross-domain wireless sensing, the privacy protection capability of cross-domain heterogeneous data is effectively improved.

The application discloses an engineering machinery activity recognition method and system based on random convolution kernel time sequence feature transformation, generates an original data txt file after collecting time sequence data from a sensor, converts the original data txt file into a csv file after pretreatment, and then utilizes a random convolution kernel transformation algorithm to perform feature transformation on the time sequence data, and sends the transformed data into a linear classifier for classification and recognition. The application uses the method of performing feature transformation on the time sequence data by using the random convolution kernel, avoids huge calculation overheads during training of a deep convolution network, fully extracts multi-dimensional time sequence features through a large number of random convolution kernels, and improves the accuracy of engineering machinery activity recognition.

The invention discloses a radar target high-resolution one-dimensional range profile recognition method, and relates to the field of radar signal processing. Comprising the following steps: S1, constructing a target high-resolution one-dimensional range profile training sample set X, a label matrix H and a label block diagonal matrix Q; s2, extracting features from the training sample set X and the tag block diagonal matrix Q, and optimizing to obtain a feature dictionary matrix D and a feature representation matrix Z; s3, constructing a linear classifier C by using the feature representation matrix Z and the label matrix H; s4, constructing a feature representation matrix Zt of the test sample set Xt according to the feature dictionary matrix D; and S5, performing target classification by using the feature representation matrix Zt of the test sample set Xt and the linear classifier C. The problem that a traditional target recognition method is poor in training effect is solved.

The application discloses a hyperspectral image classification method and system based on a physical information neural network, first acquires a hyperspectral image; constructs a spectrum-image hybrid encoder and a continuous coordinate mapping network, which are used for extracting a data-driven feature vector and a physical feature vector of the hyperspectral image; and trains and optimizes the spectrum-image hybrid encoder and the continuous coordinate mapping network; finally, inputting the hyperspectral image to be classified into the trained and optimized spectrum-image hybrid encoder and continuous coordinate mapping network, the data-driven feature vector and the physical feature vector are extracted respectively, and after the data-driven feature vector and the physical feature vector are fused, each category probability is output through a linear classifier, and the category with the maximum probability is taken as the final classification result, solving the problems that the existing classification methods ignore spectral physical prior, have poor interpretability, and have insufficient generalization ability when labeled samples are scarce.

Methods for determining black-box representations of machine learning models when information pertaining to internal states or parameters of the models are not accessible are disclosed. By using outputs of the model instead of internal states, the black-box representation is model-agnostic and provides a reliable and robust representation of the model using an external lens. The black-box representation is generated using responses from the model to a series of initialization and elicitation questions that quantify the confidence that the model has in answers it just returned. The black-box representation is then used as a training dataset for a linear classifier in order to learn performance metrics about the model.

The invention discloses a few-sample underwater acoustic target recognition method based on weighted mixing generation and adversarial domain adaptation. The method comprises the following steps: step 1, carrying out category balance operation based on weighted mixing on an original signal; 2, extracting a Mel power spectrum of the underwater acoustic signal through short-time Fourier transform and Mel filtering; step 3, performing further enhancement operations such as masking on the obtained Mel power spectrum so as to improve the generalization ability; 4, performing feature extraction on information in the frequency spectrum by using an encoder based on a convolutional neural network; step 5; classifying the extracted features by using a linear classifier, and calculating loss based on a classification result for training; step 6, performing feature extraction on the unlabeled data and the training data of the target domain by using the pre-trained convolutional neural network; step 7, trying to distinguish data sources by using a discriminator realized based on a multi-layer perceptron; 8, calculating the loss of an output result of the discriminator, and carrying out the adversarial domain adaptation fine tuning of the pre-trained convolutional neural network according to the result; 9, carrying out the test verification through employing the fine-tuned convolution to cooperate with a pre-trained linear classifier; and step 10, storing the trained model and performing deployment.

The invention relates to a dynamic gesture recognition method based on improvement, and belongs to the technical field of action recognition. The method comprises the following steps: acquiring dynamic gesture video data, preprocessing the dynamic gesture video data, inputting a dynamic gesture video sequence into a feature extraction network to extract single-frame features, splicing to obtain sequence features, processing the extracted sequence features by adopting a DSA attention mechanism, highlighting a gesture dynamic related region, and obtaining a gesture dynamic related region. The processed features are input into a coding layer for spatial embedding and position coding; a feature pyramid comprising a plurality of improvements is established, and an SRA attention mechanism is adopted in each feature pyramid to sequentially reduce the dimensionality of output features; and performing average pooling and linear classifier processing on the obtained multi-scale features, and outputting final gesture-like probability distribution. According to the method, extraction of multi-scale features of the hand region is increased, details and contexts of different levels in the gesture are captured, and therefore the recognition accuracy is enhanced.

This invention discloses a classification model for perinatal brain ultrasound images based on the USMamba framework. The classification model includes: a hierarchical visual backbone network constructed using a state-space model, consisting of four stages. Each stage contains multiple LGMB modules stacked as basic building blocks. Each LGMB module includes a Local branch and a Global branch. The Local branch performs multi-directional local feature extraction, and the outputs of the two branches are fused along the feature dimension to obtain deep features. A CASA module preprocesses the input ultrasound image using clustering adaptive convolution and sparse attention mechanisms to generate local enhancement features with structural priors. These local enhancement features are input into the backbone network, where they are used to extract deep features. Global average pooling and a linear classifier are then used to process these deep features to obtain the final classification result. Experiments demonstrate that USMamba is robust and efficient in processing complex medical images.