92 results about "Feature adaptation" patented technology

Systems and techniques are described herein for adapting a pretrained machine learning model. For instance, a process can include encoding a training image into a first feature vector, the training image including a first object located at a first location; generating a second feature vector based on a set of sinusoidal functions using a set of weights; combining the first feature vector with a second feature vector to generate a combined feature vector; processing the combined feature vector using a visual language model to obtain a second location for the first object; and adjusting the set of weights based on a comparison between the first location and the second location.

The embodiment of the invention provides a content recommendation method and device and a double-tower content recommendation model training method and device, and the content recommendation method comprises the steps: obtaining the content information of to-be-recommended content and the user information of a target user in response to a content recommendation task; the user information and the content information are input into a double-tower content recommendation model, recommended content for the target user is obtained, the double-tower content recommendation model comprises a user tower and a content tower, the user tower is used for extracting user features based on the user information, and the content tower comprises a feature adaptation layer; the feature adaptation layer is used for obtaining target content features corresponding to a task target of the content recommendation task based on the content information, and the recommendation content is obtained by decoding based on the target content features and the user features. The content representation can be dynamically adjusted according to the task target through the feature adaptation layer, so that the same content presents differentiated feature expression under different tasks, and the adaptation precision of the recommendation result to the specific task target is improved on the premise of not changing the double-tower structure.

The invention discloses an ultra-high-definition video stream adaptive coding method based on deep learning visual saliency, and the method comprises the steps: carrying out the five-scale Gaussian filtering processing and image pyramid construction of a video frame, and combining Sobel gradient, Laplacian edge and local binary pattern feature extraction to generate a multi-scale feature map; a pre-training saliency detection network is adopted, and a smooth saliency thermodynamic diagram is generated through processing of a feature adaptation layer, a residual encoder, a self-attention mechanism and a transposed convolution decoder; dividing the video frame into a high region, a middle region and a low region according to the saliency thermodynamic diagram, and establishing a regionalization coding parameter table; performing differentiated prediction modes, motion estimation and quantization strategies on different salient regions; and organizing coded data according to an H.265 / HEVC standard, and embedding the saliency thermodynamic diagram into supplementary enhancement information for transmission. According to the method, the important region concerned by the user can be intelligently identified, a differentiated coding strategy based on content semantics is realized, and the coding efficiency is remarkably improved.

The invention provides a weak password detection method based on multi-modal feature fusion and dynamic behavior analysis. According to the method, multi-dimensional feature extraction is innovatively introduced, including password entropy, semantic relevance, user historical behaviors, system login frequency and the like, so that user feature adaptation is improved, and dependency on a static dictionary is reduced; in addition, through real-time interactive feedback, instant pushing of password strength evaluation and safety suggestions can be realized. Therefore, by means of the method, the problems that in an existing weak password detection technology, the static dictionary dependency is high, the user feature adaptation is insufficient, and real-time feedback is lacked can be solved, the weak password detection precision and defense efficiency of colleges and universities are remarkably improved, and an efficient and easy-to-deploy password security solution is provided for the education industry.

The invention discloses a few-sample industrial processing anomaly detection method based on a pre-training model CLIP, and aims to solve detection pain points of scarcity of abnormal samples, high labeling cost and insufficient generalization in an industrial scene. The method comprises the following steps: data preprocessing: collecting an industrial image, dividing the industrial image into a support set only containing normal samples and a query set containing normal and abnormal samples, and combining data enhancement and local module clustering decomposition; cross-modal feature adaptation is carried out, text semantic anchor points are constructed through combination prompt integration, a residual adapter is inserted to optimize vision-text feature alignment, and multi-scale features are aggregated by adopting a harmonic average method; few-sample reference learning is carried out, a normal mode reference memory bank is constructed based on a prototype network, and feature learning is reinforced in combination with feature registration and alternate learning; and abnormal judgment: realizing abnormal recognition through cosine similarity calculation and a dynamic threshold value, and outputting an image-level confidence coefficient and a pixel-level segmentation image.

The invention provides a mine post personnel behavior identification method and system based on artificial intelligence, and the method comprises the steps: constructing a mine post dynamic behavior logic reference model, combing post personnel standard operation behavior and process node association logic, and dynamically adjusting the weight. Acquiring personnel behavior acquisition data, and performing adaptive processing on the data and the model to obtain to-be-identified behavior data; and calling a pre-trained AI behavior difference adaptation model to carry out bidirectional feature adaptation, and generating dynamically optimized behavior feature adaptation parameters. And based on the parameters, performing enhancement processing on the to-be-identified behavior data, mining hidden logic association, and generating an enhanced behavior difference feature set. And performing compliance verification according to a preset rule base to obtain a compliance verification result. And finally, generating a post behavior control instruction according to the result and sending the post behavior control instruction to a mine operation control terminal. According to the invention, behaviors of mine post personnel can be accurately identified, and operation safety and high efficiency are guaranteed.

The invention discloses a hyperspectral image segmentation method and device based on fusion point prompt and Markov diffusion, and relates to the technical field of hyperspectral image processing. The method comprises the following steps: performing spectrum-space dimension reduction processing according to hyperspectral initial data; on the basis of preset uniform distribution points, according to the spatial partitioning features, a spectrum-spatial feature adaptation module is used to carry out point prompt guided coarse segmentation; based on a cross double-attention mechanism, using a multi-modal fusion module to perform text-image feature fusion; performing multi-scale feature extraction by using a U-net encoder according to the hyperspectral initial data; diffusion reconstruction is carried out based on a symmetric codec convolutional network of a Markov diffusion model, and de-noised hyperspectral features and high-order fusion masks are extracted; and based on a cross entropy loss function, performing model optimization according to the segmentation prediction data. The hyperspectral image segmentation method is based on the text semantic features, fully considers the characteristics of the hyperspectral image, and is high in efficiency and robustness.

The application discloses a kind of multi-text feature adapter enhanced professional accomplishment named entity recognition method.The application is according to the meaning of content to professional accomplishment named entity and is labeled, and based on "BIO" method, professional accomplishment named entity is labeled with character tag;And with the characteristics and advantages of BERT model in the field of natural language processing, multi-text feature adapter is integrated into BERT to fine-tune the model, and MFEBERT model is proposed;Professional accomplishment named entity recognition model of MFEBERT+BiLSTM+CRF is constructed, MFEBERT utilizes the multi-text feature adapter to fuse the character-level features, lexical-level and part-of-speech-level fusion features of professional accomplishment named entity, learns the constraint conditions of professional accomplishment named entity by BiLSTM+CRF, and finally realizes the intelligent identification of professional accomplishment named entity, provides important technical support for scientific, accurate and effective construction of professional accomplishment evaluation index system, and helps to promote the innovative construction and application of education evaluation system.

The application relates to the technical field of automatic driving perception, in particular to a space prior calibration and feature adaptation method and system for a sparse perception architecture, which freezes the interpolated position embedding matrix into a non-trainable state, so that the space layout sensitive representation (such as object relative position and scale relationship) learned in the pre-training stage is completely preserved, the space prior destruction caused by the re-learning of a downstream task is avoided, and the overfitting risk is effectively reduced. A learnable 1x1 one-dimensional convolution is used for calibration along the channel dimension, so that the trainable parameter quantity and the input resolution are completely decoupled. The image features output after the space prior protection and the channel response calibration have accurate geometric structures, so that the Deformable Attention (deformable attention) mechanism in a sparse perception method such as Sparse4D is accurate in positioning and rich in feature semantics when four-dimensional key point sampling is performed, and the overall performance of three-dimensional target detection is improved.

The application provides a kind of operation and maintenance scene visualization report generation method and system based on natural language question and answer, belongs to intelligent operation and maintenance technical field, method includes: receiving the natural language query input by user;Through the pre-training of large language model and the preset operation and maintenance terminology dictionary, the natural language query is parsed and entity is extracted, and the entity-field association table containing demand type is generated;According to demand type and semantic feature, the query type is judged to be data query or visual query;Based on the judgment result, generate structured query language sentence or visual instruction;Query is executed to the interface business database to obtain raw data;Raw data is processed and analyzed to obtain analysis result data;Call data feature adaptation algorithm to match chart type, generate and output visual report.The application realizes the full-link automation from natural language input to visual report output, reduces the operation and maintenance data interaction threshold, improves the operation and maintenance data processing efficiency and accuracy.

The invention discloses a few-sample anomaly detection method based on two-stage training, relates to the technical field of few-sample anomaly detection, and comprises a two-stage training network (TSTNet) based on metric learning, and in a pre-training stage, geometric consistency loss is introduced to promote convergence of normal samples and geometric variants thereof in a feature space. In a metric learning stage, an edge feature adaptation network is designed, fuzzy normal features can be clustered adaptively, and abnormal features are effectively isolated. A large number of experiments on an MVTecAD data set and a VisA data set show that the scheme exceeds the existing level, and under the two-sample experiment setting, the performance of the method on image level AUROC is improved by 1.2% and 2.4% respectively compared with the MVTecAD data set and the VisA data set.

The application discloses a video target recognition method based on multi-model hot switching, and belongs to the technical field of computer vision and video processing. The method comprises an arbitration module, a switching control module and a feature adaptation and buffer module. The arbitration module generates a switching preparation signal by extracting multi-dimensional indexes such as optical flow mean, local variance, target density and scene confidence in real time through a lightweight channel independent of main reasoning. The switching control module performs atomic replacement of a computation graph pointer in a vertical blanking period, and realizes millisecond-level hot switching with zero frame loss by combining an asynchronous pre-copy and a chasing mechanism of a double buffer. The feature adaptation and buffer module solves tensor shape mismatch between heterogeneous models through a pre-compiled adaptation layer. The application also provides optimization schemes such as multi-index nonlinear fusion decision, zero-copy memory management, local slice focus reasoning and edge-cloud hierarchical unloading, significantly reduces switching delay, guarantees continuous recognition of a video stream, and is suitable for edge computing scenes with limited resources.

The invention relates to a visual feature adaptation method based on flight scene category driving, and belongs to the technical field of image processing. The method comprises the following steps: establishing a scene classification data set with labels; constructing an image feature extraction algorithm set; performing feature extraction on the images of the various flight scenes in the scene classification data set by using algorithms in the algorithm set one by one, and performing feature merging and dimension reduction to obtain representative features corresponding to the various scene images; then image clustering is carried out, and representative features of all images under each algorithm are clustered into clusters with the same number of scene categories through clustering; and according to the clustering result and the actual scene classification label of the image, calculating the index of the feature adaptation degree, which is used for representing the adaptation degree of the image feature extraction algorithm for processing a certain scene image. According to the method, accurate comparison of feature adaptation degrees of different feature extraction algorithms under various flight scene categories is realized, and a scientific basis is provided for feature selection in an aircraft visual task.

The invention discloses a scene feature joint modeling method, system, equipment and medium, and relates to the technical field of power grid engineering, and the method comprises the steps of completing spatial-temporal scale calibration, extracting independent feature vectors, generating a fusion feature map, screening key influence factors, generating comprehensive feature representation, and measuring and calculating cost and engineering quantity. The system comprises an acquisition and alignment module, a feature extraction module, a feature fusion module, a correlation analysis module, a depth modeling module and a prediction module. By constructing a space-time unified reference framework, multivariate cross-modal data is mapped to a unified space-time scale, and the problem of data islands is solved; a multi-resolution feature adaptation strategy is utilized to extract fusion features considering shallow details and deep semantics, and feature extraction comprehensiveness is ensured; and the influence of the key factors is quantified by establishing a regression mapping model, so that the accuracy and reliability of power grid project cost measurement and calculation are greatly improved.

This application discloses a method and apparatus for scientific chart data reconstruction based on semantic understanding and feature adaptation, relating to the field of scientific chart data reconstruction technology. The method includes: acquiring an image of the scientific chart to be reconstructed and identifying the data representation area and coordinate frame area; performing semantic parsing on the scale numbers within the coordinate frame area and constructing a mapping relationship between image pixel coordinates and image scale coordinates; identifying data markers within the data representation area and obtaining data sequences with different marker styles through multimodal feature vector clustering; applying the mapping relationship based on the image pixel coordinates of the data markers in each data sequence to obtain the scale coordinates corresponding to each data marker, thereby performing scientific chart reconstruction to obtain a reconstructed scientific chart image. This application solves the fundamental problem that existing tools are limited in scope due to their reliance on predefined templates, enabling a single technique to cover the vast majority of chart types in scientific publications.

The invention discloses a necking detection method based on geometric feature self-adaption. The necking detection method comprises the following steps that S1, the longest main axis of a simulation graph is determined; s2, selecting N feature points on the longest main shaft; s3, obtaining the minimum key size of the simulation graph at each feature point; and S4, judging whether the simulation graph has a necking risk or not based on each minimum key size. By identifying the geometric centroid and the longest principal axis of a graph, adaptively selecting feature points and performing omnidirectional scanning, the measured minimum critical dimension is compared with a set safety threshold, and a risk point which is most likely to generate necking on the graph is accurately positioned, so that the automation and objectification of risk detection are realized, and the risk detection efficiency is improved. Subjectivity and experience dependence of manual marking are avoided, and consistency of detection results is guaranteed; meanwhile, the method has good adaptability to complex or deformed contours, missing detection and misjudgment are effectively avoided, and the yield and reliability of chip manufacturing are improved.

The present application relates to the technical field of speech recognition, and discloses a psychological counseling real-time speech recognition method based on multi-modal data, comprising: constructing a gender-specific speech psychological feature mapping baseline, analyzing and judging the emotional stability tendency degree, and combining the amplitude peak value proportion to judge the expression tendency degree; dynamically correcting the fundamental frequency mean amplitude to solve the psychological feature misjudgment caused by individual pronunciation difference; detecting the stress position, key words, speech speed and pause features through the fundamental frequency mutation and amplitude mutation, constructing a multi-dimensional emotion analysis model, and optimizing the emotion focus positioning, emotion type judgment and change trend tracking problems; designing a comprehensive stability value calculation method, simultaneously reflecting the emotional stability and the influence degree of key words, and providing a psychological health evaluation quantitative index; constructing a three-level processing mechanism, and through the fusion of the exclusive baseline and multi-modal evidence, performing feature adaptation, cross-validation and decision correction.

This invention discloses a multimodal visual inspection method for cross-scale parts based on a CAV model, comprising the following steps: S1, constructing a cross-scale annotation system; S2, constructing a language-guided improved CAV model: the improved CAV model includes a feature extraction module, a multimodal fusion module, a CAV concept learning module, a cross-scale feature adaptation module, and a detection decision module; S3, model training: training the improved CAV model using the multimodal dataset constructed in step S1, generating CAV vectors of part detection-related concepts through language guidance, optimizing model parameters, enabling the model to accurately capture the multimodal features and defect features of cross-scale parts; S4, data acquisition and preprocessing; S5, feature fusion. This invention significantly improves detection accuracy and achieves accurate identification of cross-scale defects.

The embodiment of the application discloses a domain generalization target detection method based on a large model guide and related equipment, a training image set with the same content is obtained, and a foreground enhanced image set is generated in combination with foreground mask guidance, which not only helps to enrich the diversity of data enhancement from the root by means of multiple training data, but also focuses on the target foreground for feature enhancement through the foreground mask, thereby effectively retaining the positioning characteristics of the target while improving the richness of data distribution and avoiding damage to the positioning accuracy of the target. The style of the foreground enhanced image set and the to-be-predicted image is fused to realize feature adaptation, and then a hyper-domain invariant feature encoding is used to obtain a feature map after feature encoding. On the basis of retaining key feature information, the feature map after feature encoding efficiently extracts domain invariant features with strong generalization, thereby avoiding feature information loss and improving feature representation capability, so that accurate target detection results are obtained.

This invention relates to the field of remote sensing image processing, specifically to a large-model adaptation and fine-tuning method for remote sensing land cover classification based on Fourier transform. It combines Fourier transform with multi-feature adaptation and fine-tuning of SAM (Simulated Aspect-Oriented Mapping), constructing a SAM remote sensing land cover classification model coupled with Fourier transform and multi-feature adaptation and fine-tuning. The method includes the following steps: data acquisition and preprocessing, construction of a weight adaptation mechanism, Fourier transform module design, construction of a multi-feature adapter, model structure design, model training, and model performance evaluation. This invention significantly improves the model's ability to represent complex scenes and multi-scale features, effectively adapting to the differences in features at different scales and the complexity of scenes in remote sensing images. Furthermore, by fully utilizing the pre-training capabilities of SAM, the model can adapt to various remote sensing image resolutions and input sizes, greatly improving its applicability and scalability, and providing an efficient and practical new method for remote sensing land cover classification.

The application provides a kind of application software development compatibility debugging method and device for heterogeneous equipment, and is related to application software debugging.The application constructs equipment dynamic characteristic map by collecting and integrating equipment hardware, system, resource, interface and running state characteristics through multi-modal feature fusion technology.Secondly, the real-time iterative optimization of compatibility calculation weight is realized by introducing reinforcement learning algorithm, and the feature adaptation accuracy is improved by combining the improved normalization model.Through abnormal fingerprint extraction and tracing algorithm, the precise positioning and root cause analysis of incompatibility problem are realized.Finally, based on dynamic confidence threshold, hierarchical adaptation, intelligent repair and cross-device debugging are executed.The application breaks through the limitations of traditional fixed strategy debugging, realizes the intelligentization, self-adaptation and precision of compatibility debugging, greatly improves the debugging efficiency and reduces the adaptation cost, and is suitable for various heterogeneous equipment platforms.

The invention relates to an image tampering positioning method and device, equipment and a storage medium, and the method comprises the steps: extracting two types of initial features from an input image, namely a spatial feature and a high-frequency feature; respectively sending the two types of features into a pre-trained visual backbone network with frozen parameters, introducing two groups of learnable prompt word vectors in front of each layer of the visual backbone network to respectively act on a spatial branch and a high-frequency branch, and carrying out multi-view feature adaptation and enhancement; s3, feature alignment and fusion: in the alignment stage, respectively calculating cross-branch channel attention and space attention and carrying out interactive weighting, and in the fusion stage, applying cross-branch deformable attention by adopting multi-expansion-rate cavity convolution; and S4, sending the multi-scale features output by the visual backbone network into a tampering region positioning decoder to generate a tampering mask. Compared with the prior art, the method has the advantages that the prior knowledge of the pre-training model can be fully mobilized under the condition that the training data is limited, and the higher tampering positioning capability is obtained.

The invention relates to the technical field of artificial intelligence monitoring, in particular to a continuous monitoring method and system for multi-modal physiological parameter fusion, and the method comprises the steps: carrying out the continuous collection of physiological electric signals of a target individual in a plurality of behavior states; performing period segmentation on the multi-lead electric signal according to the heartbeat rhythm, and standardizing the multi-lead electric signal into characteristic fragments with the same time length and channel number; performing two-stage weighting processing on each feature fragment to obtain high-weight feature representation; performing feature alignment on the high-weight feature representation according to the corresponding behavior state, and executing fusion operation; and inputting the fusion feature representation into a convolutional neural network, performing individual physiological state recognition, and periodically judging whether the target individual has a physiological abnormality risk at present. According to the method and the device, the problem of signal distortion caused by neglecting multi-behavior state switching in the prior art is effectively solved, the feature adaptability and the cross-state recognition robustness in a dynamic scene are remarkably improved, and continuous and accurate physiological anomaly risk tracking is realized.

Systems and techniques for adapting pre-trained machine learning models are described herein. For example, a method can include encoding a training image as a first feature vector, the training image including a first object located at a first position; generating a second feature vector based on a set of sinusoidal functions using a set of weights; combining the first feature vector with the second feature vector to generate a combined feature vector; processing the combined feature vector using a visual language model to obtain a second position of the first object; and adjusting the set of weights based on a comparison between the first position and the second position.

The invention belongs to the technical field of deep learning and image recognition, and discloses an image semantic segmentation method based on a feature complementation fusion model, the method is used for road disease recognition, the feature complementation fusion model comprises an encoder and a decoder, the encoder comprises a TransformerBlock module and a CNNBlock module, and the decoder comprises a CNNBlock module. The decoder comprises an upsampling Upsample Block module, a multi-scale feature fusion module, an output module and a context information recovery module. The method comprises the steps of constructing a road disease data set, preprocessing the road disease data set, labeling the data set, and constructing a feature complementary fusion model. Inputting the training set data into the feature complementation fusion model for training; and evaluating model performance by using test set data. According to the method, the problems of insufficient multi-scale feature adaptation and fuzzy boundary are solved.

The invention provides a cross-game user LTV curve prediction method based on a dynamic quantile correction mechanism. The cross-game user LTV curve prediction method comprises the following steps: (1) obtaining basic training data and processing feature engineering; (2) quantile random forest basic model set training; (3) obtaining short-term alignment data of a target game and carrying out feature adaptation processing; (4) dynamically calculating an optimal quantile parameter; (5) predicting the long-term life cycle value of the new user; and (6) fitting a long-term life cycle value recovery curve. The core idea of the invention is as follows: a traditional point prediction model is transformed into a quantile prediction model, a dynamically optimized quantile parameter calculation module is introduced, and the module calculates a group of optimal quantile parameters by utilizing short-term data (such as 60 / 90 days) which is relatively easily obtained by a target game (or a new game), so that the optimal quantile parameters are obtained. The method is used for correcting the predictive output of the basic model for long-term targets (such as 180 days), thereby adapting to data drift and cross-game differences.

The application discloses a kind of based on incremental expert network continuous learning mechanism medical image analysis method, system, electronic equipment, storable medium, it is related to deep learning and medical image analysis technical field.MRI image with label from multiple medical image domains is acquired;Medical image analysis network model is constructed, corresponding domain incremental expansion module is introduced in shared feature adaptation module, and the feature prototype representation of each domain is constructed and updated simultaneously by domain selection control module;Medical image analysis network model is trained based on MRI image;The medical image to be segmented is input into medical image analysis network model, and the similarity of its feature and each domain feature prototype is calculated by domain selection control module to determine matching domain, corresponding domain incremental expansion module is called to participate in feature processing, after being processed by feature extraction, feature adaptation and prediction module, the segmentation result of target region is generated.The application can improve the accurate segmentation effect of organ structure and lesion region.

The application provides a hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis, comprising a self-step prototype prompt generator, a context-aware prompt retriever and a prompt-guided refinement and adaptation module. In this way, a ProtoRAP framework is proposed to realize rolling bearing cross-domain fault diagnosis through prototype-guided knowledge encoding, context-aware dynamic retrieval and gated-driven feature adaptation. The hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis provided in the embodiment are significantly superior to existing advanced methods in terms of diagnosis accuracy and cross-domain stability in the variable working condition diagnosis task of multiple rolling bearing fault data sets, verifying the effectiveness and robustness of the hybrid prompt learning network and system for rolling bearing cross-domain fault diagnosis in complex industrial scenarios.

The invention relates to the technical field of speech recognition, and discloses a psychological counseling real-time speech recognition method based on multi-modal data, which comprises the following steps: analyzing and judging an emotion stability tendency degree by constructing a gender-exclusive speech psychological feature mapping baseline, and judging an expression tendency degree in combination with an amplitude peak value proportion; the fundamental frequency mean amplitude is dynamically corrected, and psychological feature misjudgment caused by individual pronunciation difference is solved; detecting accent positions, keywords, speech speed and pause features through fundamental frequency abrupt change and amplitude abrupt change, constructing a multi-dimensional emotion analysis model, and optimizing emotion focus positioning, emotion type judgment and change trend tracking problems; designing a comprehensive stable value calculation method, reflecting the emotional stability and the keyword influence degree at the same time, and providing quantitative indexes for mental health assessment; and constructing a three-level processing mechanism, and performing feature adaptation, cross validation and decision correction through exclusive baseline and multi-modal evidence fusion.

The application provides an out-of-focus image estimation method based on region feature adaptation and attention fusion, which uses a common Encoder-Decoder structure as a basic network for an image restoration task. Secondly, a skip connection is added in the basic network, and a channel attention mechanism is introduced in the decoder, so that the decoder can better retain and utilize the feature information extracted in the encoder. At the same time, a region feature adaptation network is used for image-to-image regression, which is attached to the basic network to minimize the domain difference between the synthesized features and the real features, and a channel attention mechanism is also introduced in the region feature adaptation network to further enhance the representation ability of the network. Finally, during training, the region feature adaptation network is first trained as a discriminator, and the discriminator loss is used to classify the features of the synthesized domain and the real domain, and then the basic network is trained to minimize the domain difference between the synthesized and real out-of-focus image features.