33 results about "Feature transform" patented technology

A multimodal entity linking method based on dual encoders and a hybrid expert mechanism is proposed. This invention relates to multimodal entity linking technology at the intersection of natural language processing and computer vision. Addressing the problems of low inference efficiency, insufficient cross-modal interaction, and shallow modal fusion in existing methods, this invention proposes a multimodal entity linking method based on dual encoders and a hybrid expert mechanism. A dual-tower architecture is used to independently encode mentions and entities. Entity embeddings can be pre-computed offline and indexed, and linking is completed during inference through fast vector retrieval. A hybrid expert mechanism is introduced to achieve adaptive feature transformation of samples, and a gating network dynamically selects expert combinations. Bidirectional cross-modal attention is used to establish fine-grained alignment at the word-image block granularity. A channel attention mechanism dynamically balances the contributions of textual and visual modalities. The model is jointly optimized by multiple constraints, including load balancing loss. This invention achieves efficient retrieval while maintaining deep inference capabilities, simplifies inference time complexity, and is suitable for scenarios such as knowledge graph construction and intelligent question answering systems.

This application provides an image compression system, an image processing method, an encoding / decoding method, and an electronic device. The image compression system includes a first selection module, an entropy encoding module, an entropy decoding module, a quantization module, N encoding networks, and one decoding network. The N encoding networks have different encoding losses. The first selection module is used to select a target encoding network from the N encoding networks based on the number of times the image to be encoded has been encoded. The target encoding network is used to perform feature transformation on the image to be encoded to obtain a first feature map. The quantization module is used to quantize the first feature map to obtain a second feature map. The entropy encoding module is used to entropy encode the second feature map to obtain a bitstream. The entropy decoding module is used to entropy decode the bitstream to obtain a third feature map. The decoding network is used to perform feature transformation based on the third feature map to obtain a reconstructed image. This effectively reduces the loss ratio of the image after multiple encoding and decoding operations.

The present application relates to the technical field of speech recognition, and provides a speech recognition method and device, electronic equipment and storage medium, wherein the method comprises: performing feature extraction on a speech signal to be recognized, inputting an extracted acoustic feature sequence into a speech recognition model to obtain target recognition text that has been optimized by text; wherein the speech recognition model comprises an encoder and a decoder, and a hybrid expert module for a text optimization task is embedded in the encoder; the encoder is used for performing layer-by-layer encoding processing on the acoustic feature sequence, and performing feature transformation corresponding to the text optimization task on intermediate level features through the hybrid expert module in the encoding process to obtain encoding features containing text optimization semantics; and the decoder is used for decoding the encoding features to obtain the target recognition text, effectively solving the problems of bloated architecture and error accumulation caused by separation of recognition and text optimization in a traditional speech recognition system, and achieving dual improvement of inference efficiency and recognition quality.

The application relates to a partial discharge diagnosis method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring an acoustic-electric signal to be analyzed and task constraint information for providing constraint information, performing feature coding processing on the acoustic-electric signal to be analyzed to obtain an acoustic-electric coupling feature, performing feature transformation on the acoustic-electric coupling feature and aligning the acoustic-electric coupling feature to a preset state parameter domain to obtain a state analysis feature, then performing acoustic-electric joint diagnosis reasoning based on the task constraint information and the state analysis feature to obtain a partial discharge reasoning feature, and performing semantic generation based on the partial discharge reasoning feature to obtain a diagnosis result. The method can solve the acoustic-electric cross-modal feature space dislocation problem, significantly improve the precision and robustness of partial discharge diagnosis under complex working conditions, can synchronously output accurate partial discharge state description and fault labels, and significantly improves the accuracy and intelligent analysis depth of partial discharge diagnosis.

This application relates to the field of video surveillance and image processing technology in coal mines, and discloses a method and apparatus for panoramic stitching of working faces using an improved optical flow algorithm. The method includes the following steps: First, multi-scale adaptive preprocessing is performed on the original image; an anti-interference weight model based on texture entropy and gradient is constructed, and a sparse optical flow field is generated using Retinex illumination component separation technology to suppress dust and illumination interference. Second, feature matching is guided by the sparse optical flow field, and a dynamic region mask is generated by combining target detection. The optical flow transformation and feature transformation matrices are then adaptively weighted and fused. Finally, an improved Poisson fusion algorithm is used to repair dynamically occluded areas, and the optimal seam path is planned by calculating the optical flow variance. A dynamic window is then used for smooth fusion. This invention effectively solves the problems of stitching misalignment and artifacts caused by high dust levels, illumination fluctuations, and rapid equipment movement in coal mines, achieving high-quality seamless panoramic stitching.

The present application belongs to the technical field of sea clutter prediction, and particularly relates to an adaptive Fourier and Mamba sea clutter prediction method, system and medium, which comprises the following steps: obtaining original radar echo data, and performing input preprocessing of time-space enhancement; performing frequency domain transformation through Fourier, introducing learnable frequency band parameters, adaptively separating frequency bands, performing feature transformation and enhancement on the separated frequency bands, and obtaining time domain features through inverse Fourier transformation; the time domain features are activated to enter two parallel paths, are processed through Mamba blocks, and are fused through cross modulation and enhancement to obtain a final feature tensor; after layer normalization, the feature tensor extracts global context representation, and a multilayer perception machine is used to map the converged context representation to amplitude prediction values of future k time steps; the stability of prediction and the adaptability to sudden events are significantly improved, and prediction bias is overcome.

The application discloses a kind of asynchronous distributed radar detection positioning integration method based on multidimensional feature transformation, each independent operation asynchronous radar carries out local low threshold detection, only the point track that passes threshold is transmitted to fusion center, reduce the calculation burden of fusion center under the premise of not sacrificing too much fusion detection performance;Fusion center is mapped to feature space by multidimensional extended feature transformation proposed, and multidimensional feature space is quantized sampling by space delimitation and importance sampling operation;Iterative detection is carried out to multidimensional feature space by two-stage detection method, and detection point track and initial track are output;Based on extended speed feature transformation, spatial velocity estimation is obtained using Doppler information and angle measurement information, and positioning estimation is estimated by combining spatial velocity estimation and position measurement.The application improves the detection performance and real-time processing performance of traditional batch processing algorithm under high false alarm rate by optimizing the acquisition method of accumulation score table and detection process.

The application provides a random subspace-based coastal zone ecological disaster mitigation assessment method, device, equipment and storage medium, a sample matrix is constructed, and a feature transformation matrix, a discrete coding matrix and a clustering center matrix are initialized, in an iterative updating process, a similarity matrix of each sample belonging to a feature transformation matrix array vector set is calculated, different types of coastal zone samples can be associated to different feature subspaces, and therefore the limitation of a single subspace assumption is broken through; finally, key risk assessment indexes are screened out according to L2 norm ordering of each row vector of the feature transformation matrix, and targeted ecological disaster mitigation risk assessment on a coastal zone region with spatial heterogeneity is realized.

This invention relates to the field of image communication technology, specifically disclosing an image semantic communication method and system based on a semantic representation recalibration network. To improve the effectiveness of intermediate semantic representations under low bit rate and complex channel conditions, this invention introduces a semantic representation recalibration network during the deep feature transformation process at the encoding end. This network constructs two semantic representation recalibration processes within the feedforward network. It extracts local context information under different receptive fields through a multi-scale local context extraction branch, and combines channel attention and spatial attention branches to adaptively enhance key features in the intermediate semantic representation and suppress redundant features, thereby improving the effective semantic carrying capacity under a limited transmission budget. This effectively enhances the effective semantic carrying capacity within a limited transmission budget under bandwidth-constrained and noisy channel conditions, alleviates the information loss problem in low-dimensional transmission representation, and ultimately achieves an effective improvement in image reconstruction quality.

This invention discloses a ground-to-air pedestrian re-identification method, system, device, and storage medium, which correspond to a crime scene. The solution employs a learnable feature projection mechanism to explicitly model and bridge the significant geometric and appearance gap between aerial and ground-based perspectives, thereby more effectively extracting identity-sensitive and perspective-sensitive feature representations while suppressing interference information introduced by extreme perspective differences. Furthermore, by collaboratively achieving cross-perspective alignment and decoupling at the levels of feature transformation, supervisory signals, and optimization objectives, the model is guided to learn the optimal cross-perspective pedestrian representation, thus maintaining high robustness and strong generalization ability in complex scenarios and accurately achieving ground-to-air pedestrian re-identification. This solution has broad application prospects in scenarios requiring integrated ground-to-air perception, such as urban three-dimensional security, emergency search and rescue, and security for large-scale events.

This invention discloses a RecycleNet2 model and data processing method based on information loss, belonging to the field of computer technology. The model includes K-1 sequentially cascaded RecycleNet2 modules, K feature transformation modules, a fully connected layer, a concatenation and fusion layer, and an output layer. The main branch of the RecycleNet2 module uses 1×3 and 3×1 asymmetric convolutions to extract preliminary features, while the side branch uses a subtraction layer to subtract the input from the preliminary feature map, extracting and recovering lost details to generate recovered feature maps. The K recovered feature maps are then subjected to secondary feature mining via 1×1 convolutions built into the feature transformation modules before being input into the corresponding fully connected layers. Finally, the concatenation and fusion layer performs unified fusion of all feature output representations, and the output layer outputs the classification result. This invention effectively solves the problem of information loss during convolution, improving the depth representation of network features and classification accuracy.

This invention discloses a data analysis method and system for the production and testing of miniature bearings based on big data, belonging to the field of data processing technology. The invention pre-trains a recursive time-series network by collecting vibration signals from bearings in known states. This network, combined with a transient noise suppression operator, obtains the evolutionary feature vectors of bearings in the same batch. Static feature vectors are extracted by combining processing time sequences and end-face images. After splicing, visually missing samples are corrected, and a reconstructed kernel feature matrix is ​​constructed. Based on this matrix and real labels, bidirectional alternating optimization is performed to obtain a feature transformation and mapping matrix, establishing a bearing state database. Multimodal data of the bearings under test are collected to generate initial feature vectors to be detected. After mapping, the similarity with database samples is calculated to output the final detection result. This invention deeply integrates multimodal features, effectively overcoming interference from workshop visual contamination, thermal expansion, and group tolerance drift, significantly improving detection accuracy and robustness.

The application relates to the technical field of chip defect identification, and discloses a wafer chip defect detection method, device, equipment and medium. In the method, a Mamba space modeling network is used, so that when a high-resolution original gray image of a wafer chip is processed, the calculation complexity is reduced from a square level of an existing method to a linear level, the inference speed is improved, and the demand of real-time detection of an industrial production line is met. The spatial propagation mechanism of the Mamba space modeling network effectively captures the spatial propagation characteristics of wafer chip defects, improves the identification accuracy and spatial consistency of systematic defects and continuous defects, the learnable spline base function of the KAN network enhances the interpretability of the feature transformation process, the decision basis of the detection result is transparent and traceable, and quality auditing and process improvement are facilitated.

This invention provides a three-dimensional spectrum prediction method, system, device, and medium based on multi-scale graph convolutional networks, belonging to the field of spectrum prediction technology. The method involves continuously acquiring received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data and constructing a feature matrix. The feature matrix is ​​transformed using a learnable parameter matrix of preset dimensions. An adaptive threshold is set based on the median score of all nodes to generate a weighted adjacency matrix. First-order and second-order graph convolution operations are performed on the weighted adjacency matrix, and feature transformations are performed using multi-dimensional learnable weight matrices to form multi-dimensional fused features. These multi-dimensional fused features are expanded according to a time series, and a gating mechanism is constructed through element-wise multiplication and residual connections. The prediction result is output through a linear transformation layer. This improves the accuracy of spectrum prediction, enabling the model to output more accurate prediction results when processing long-term spectrum data.

A combination feature screening method comprises: inputting a plurality of single-value features contained in a user sample into a first model to obtain a prediction value about user behavior, wherein the first model comprises N network layers for performing feature transformation; decomposing a first combination feature into K interaction chains, the first combination feature being composed of at least two features in the plurality of single-value features as component features, and one interaction chain being used to represent N times of feature transformation performed on the component features through the N network layers respectively, and the N times of feature transformation including feature transformation between different component features; calculating an importance score of the first combination feature based on an influence of each interaction chain in the K interaction chains on a first loss, the first loss being obtained based on the prediction value and a sample label; and screening the first combination feature based on the importance score of the first combination feature. The method realizes combination feature mining of an arbitrary order, and the time consumption of combination feature importance verification is short.

The CT image segmentation method and the electronic device are disclosed. The CT image and the structured meta information corresponding to the CT image are acquired, the CT image is subjected to feature extraction to obtain CT image features of multiple levels, the structured meta information is subjected to feature coding to obtain text features, the text features are subjected to linear transformation to obtain channel gating weights, the CT image features are subjected to channel weighting processing based on the channel gating weights to obtain channel weighted features, the channel weighted features are subjected to feature enhancement to obtain CT image enhanced features, the CT image enhanced features are subjected to decoding to obtain guide features, the guide features and the CT image features of the corresponding level are subjected to feature transformation in the spatial dimension and the channel dimension to obtain joint gating weights, the CT image features are subjected to residual fusion processing based on the joint gating weights to obtain gating output features, and the CT image segmentation is performed based on the gating output features to obtain a CT image segmentation result, thereby improving the segmentation accuracy of the CT image.

The application discloses a multi-modal image matching method based on phase consistency structure information distillation. The method is used on a pair of multi-modal images to be matched. Firstly, stable structure prior information is obtained through phase consistency calculation, and is combined with multi-scale features extracted by a convolutional neural network to form structure-guided feature expression. Then, structure distillation training is performed on the network by using structure consistency loss and cross-modal consistency loss, so that the coarse scale and fine scale features obtained have stronger structure consistency under the cross-modal condition. On this basis, a feature transformation module stacked alternately by self-attention and cross-attention is used to complete coarse scale global matching to generate a candidate matching pair. Then, the candidate matching pair is refined locally by fine scale feature matching, and a final matching set is obtained. The application deeply integrates phase consistency structure prior and a deep learning matching framework, guarantees the generalization of cross-modal matching, and improves matching precision and calculation efficiency.

Systems and methods for performing video object segmentation are provided. In an example, video data representing a sequence of image frames and video data representing an object mask can be received at a video object segmentation server. Image features can be generated based on a first image frame of the sequence of image frames, image features can be generated based on a second image frame of the sequence of image frames; and object features can be generated based on the object mask. A transformation matrix can be computed based on the image features of the first image frame and the image features of the second image frame; the transformation matrix can be applied to the object features, resulting in transformed object features. A predicted object mask associated with the second image frame can be obtained by decoding the transformed object features.

This invention discloses an underwater image enhancement method based on lightweight global context modeling and partition meta-optimization, belonging to the field of underwater image enhancement technology. The method includes: reading pre-trained enhancement network parameters θ and loading them into an LGM-Net model; when used for edge-side or real-time inference, performing structural reparameterization on the locally reproducible multi-branch representation. The underwater RGB image to be enhanced is acquired and preprocessed to obtain an input tensor. Initial features are obtained through shallow feature extraction, followed by step-by-step encoding, downsampling, and feature transformation to form multi-scale pyramid features. Projective free sparse expert global modeling is performed at the deepest bottleneck, followed by step-by-step decoding and reconstruction. The final reconstructed features are obtained by skip-connection fusion with the corresponding scale feature channels of the encoder, mapped back to the RGB space by the output reconstruction head, and normalized. The enhanced image is then output to a display or storage device. This application effectively alleviates structural artifacts and local over-enhancement problems, while also enhancing cross-domain robustness and deployment efficiency.

This invention specifically relates to an air-ground collaborative path planning system and method for Mars exploration, comprising: acquiring multi-scale images of the Martian surface through air-ground collaboration; constructing a high-precision panoramic map based on a proposed adaptive sampling affine scale-invariant feature transformation algorithm; and combining a deep learning rock segmentation algorithm and a fast travel method to achieve global obstacle avoidance and energy-optimized path planning in complex unstructured Martian terrain environments. This enhances the autonomous navigation capability in complex Martian terrain and solves the technical problems of limited environmental perception range, insufficient obstacle recognition accuracy, and poor path planning efficiency and smoothness in existing Martian ground exploration path planning technologies. It is applicable to cross-view environmental mapping, rock obstacle recognition, and global path planning for Mars UAVs and Mars ground exploration rovers in complex unstructured terrain.

To provide a feature amount conversion learning device, an authentication device, a feature amount conversion learning method, an authentication method, and a recording medium, which perform feature amount super resolution with higher accuracy.SOLUTION: A feature amount conversion learning device 200 includes: an image patch acquisition unit 201 that acquires a first image; an image reduction unit 202 that reduces the first image to a second image having a resolution lower than that of the first image; an image enlargement unit 103 that enlarges the second image to a third image having the same resolution as that of the first image; a feature amount extraction unit 104 that extracts a first feature amount as a feature amount of the first image and a second feature amount as a feature amount of the third image; a feature amount conversion unit 105 that converts the second feature amount into the third feature amount; and a learning control unit 208 that allows the feature amount conversion unit 105 to learn a feature amount conversion method on the basis of a comparison result between the first feature amount and the third feature amount.SELECTED DRAWING: Figure 3

This invention relates to the field of computer vision technology and discloses a UAV visual detection method based on an improved YOLOv5s. It addresses the problems of existing detection methods failing to meet the lightweight deployment requirements of embedded UAV platforms and the difficulty in fully capturing contextual information during lightweight deployment, thus affecting detection performance. The Stem layer of this invention enhances initial feature extraction capabilities while maintaining high resolution through a multi-branch parallel structure; the Shuffle module achieves efficient feature transformation through channel segmentation and channel shuffling; the Shuffle downsampling module implements weighted feature fusion during downsampling through an adaptive gating mechanism; the C3Ghost module utilizes Ghost convolutions to first perform a small number of standard convolutions to generate intrinsic feature maps, and then generates more feature maps through low-cost linear transformation; the NextConv module optimizes the Conv structure in PANet, effectively expanding the receptive field of the feature maps without significantly increasing computational cost.

This invention discloses a generative model-driven distributed remote sensing image encoding and decoding method. The encoding end obtains a compact latent representation containing key semantic information of the source view image through simple feature transformation. The decoding end extracts common features between the side information view image and the source view image using the side information view image, and then fully extracts the rich multi-scale semantic feature representation of the side information view image spatially using a multi-scale spatial feature mining module. This feature representation contains most of the spatial semantic information of the source view image, effectively reducing the number of bits required for compressed transmission of the source view image at the encoding end. Furthermore, this multi-scale semantic feature representation can serve as a semantic prior, assisting the diffusion model in generating more robust reconstructed feature representations, ensuring that the inverse transform module recovers a higher-fidelity reconstructed image, achieving low-bitrate, high-fidelity remote sensing image compression, suitable for distributed remote sensing image compression tasks in satellite constellation scenarios.

This invention discloses a deep learning-based method for configuring automotive stamping process parameters, comprising the following steps: Step 1: collecting raw stamping data for the automotive stamping process; Step 2: obtaining standard stamping data through data preprocessing; Step 3: performing spatiotemporal separation and feature transformation on the standard stamping data; Step 4: generating a stamping parameter prediction vector using an improved ConvMixer network; Step 5: using an improved interquartile range (IQR) method and weighted correction to perform anomaly detection and correction on the stamping parameter prediction vector, obtaining an optimized stamping parameter vector; Step 6: configuring the automotive stamping process production equipment according to the optimized stamping parameter vector; Step 7: collecting stamping feedback data and incrementally learning the improved ConvMixer network. This invention improves the accuracy and stability of automotive stamping process parameter configuration through the improved ConvMixer network and the improved IQR method.

The application discloses an underwater image enhancement method based on light-weight global context modeling and partition element optimization, and relates to the technical field of underwater image enhancement.The method comprises the following steps: reading pre-training enhancement network parameters θ and loading an LGM-Net model; when used for end-side or real-time inference, performing structure re-parameterization on a local re-parameterizable multi-branch representation; obtaining an underwater RGB image to be enhanced and pre-processing the image to obtain an input tensor; performing shallow feature extraction to obtain initial features, and then performing step-by-step encoding, down-sampling and feature transformation to form multi-scale pyramid features; performing projection free sparse expert global modeling at the deepest bottleneck; then performing step-by-step decoding reconstruction, and then performing channel jump connection fusion with corresponding scale features of the encoder to obtain final reconstruction features; mapping the final reconstruction features back to the RGB space through an output reconstruction head and normalizing the final reconstruction features; and outputting an enhanced image to a display or storage device.The application effectively alleviates the problems of structural artifacts and local over-enhancement, and has the advantages of enhanced cross-domain robustness and higher deployment efficiency.

This invention discloses a method and apparatus for detoxification of large language models based on Riemann flow matching, belonging to the field of artificial intelligence technology. It addresses the pain points of existing detoxification schemes, such as easy semantic shift and difficulty in balancing detoxification effectiveness and generation quality. This invention extracts interpretable SAE features from the intermediate layer activations of a large language model using a sparse autoencoder. After toxicity subspace determination, for the SAE features that trigger intervention, a semantic null space smoothing detoxification transformation is performed using a Riemann flow matching feature transformation operator constructed from a semantic Riemann manifold. The detoxified features are decoded and reconstructed into detoxified activations, replacing the original model input activations to guide the generation of non-toxic outputs. This invention eliminates the need for fine-tuning the model's backbone weights, achieving efficient and accurate detoxification while geometrically avoiding semantic shifts, thus balancing core semantic integrity and generation fluency. It can be applied to large language model scenarios.

The present application relates to the technical field of metagenomics and data science, and provides a metagenomics binning method and system combined with prior knowledge of a reference library, which comprises: obtaining a target sequence data set, performing feature extraction on each sequence sample in the target sequence data set, and obtaining a binning feature vector set after feature transformation; comparing the target sequence data set with a reference library to obtain a species number estimation value and a confidence degree of each sequence sample belonging to different species, and then obtaining a feasible interval of binning number and a confidence clustering center of each species, and taking the same as prior knowledge; for each feasible binning number of the feasible interval of binning number, a clustering method is used to cluster the binning feature vector set, and the optimal clustering result is selected, and the optimal clustering result is used to realize binning of the target sequence data set. Compared with the prior art, the present application solves the problems of the prior art, such as the inability of existing metagenomic binning to process unknown species sequences or insufficient binning accuracy.