114 results about "Convolution filter" patented technology

The invention discloses an ultra-weak light detection system and method based on an SPAD array and time correlation photon counting, and belongs to the technical field of ultra-weak light detection. A multi-channel SPAD array detection unit; a time-to-digital conversion module; the photon arrival time analysis unit comprises a TCSPC model, a noise modeling and suppression module and a self-adaptive convolution filter; a data compression and fitting module; a spatial denoising filter; an output unit; and a master control processing unit. Based on a multi-channel SPAD array detection unit, a time-to-digital conversion module, a TCSPC model, a noise modeling and suppression module, a self-adaptive convolution filter and the like, high-sensitivity imaging and quantitative detection of materials (such as black silicon) with extremely low reflectivity are realized. The system has the advantages of being good in real-time performance, high in time resolution, high in signal-to-noise ratio, high in on-chip processing capacity, high in integration and the like, and meanwhile the bottleneck of a traditional TCSPC method for large-scale photon data processing is solved.

Techniques for speech recognition models using modulation spectrum are disclosed herein. A modulation spectrum is generated from time series data output of an encoder layer of a speech recognition model and used as input into a decoder layer of the speech recognition model to improve accuracy of the model such as for recognizing subword units. The modulation spectrum is determined by applying a convolution filter to the output of the encoder layer of the speech recognition model. The time series data and / or the modulation spectrum can be normalized. A rectified linear unit activation function can be applied to the output of the convolution filter. The output of the encoder layer may be residually connected to the output of the rectified linear unit activation function prior to being input into the decoder layer.

The invention provides a brain age estimation method based on a dynamic fuzzy learnable brain network, and belongs to the technical field of medical image processing and artificial intelligence. According to the technical scheme, the method comprises the following steps that S1, brain nuclear magnetic resonance imaging of a subject is collected, and preprocessing and data division are carried out; s2, constructing graph structure data, and performing feature extraction and position information embedding on the data; s3, constructing a dynamic fuzzy learnable brain network model comprising a main branch and a local branch, and respectively extracting global and local connection features; s4, introducing a dynamic fuzzy multi-head self-attention module into the main branch to realize effective modeling of global features; s5, a local branch dynamically models a dependency relationship between channels through a convolution filter and a learnable graph attention module; s6, after the features of the main branches and the local branches are fused, brain age prediction is carried out through a multi-layer perceptron. According to the method, the modeling capability of the brain function connection mode is improved, and the brain age prediction task can be more effectively completed.

A model for defect recognition on a surface under inspection comprises a first supervised machine learning classifier configured for classifying image units of an image into at least defective and defect-free image units; multiple sets of convolution filters perform feature extraction, wherein each such set comprises one or more convolution filters defined by a set of individual filter parameters; said convolution filters are applied to one or more labelled reference images of a surface to obtain multiple groups of feature maps; multiple trained first machine learning classifiers are obtained using said feature maps; each trained first machine learning classifier is provided with an input test set and a respective quality parameter is obtained; an optimized set of convolution filters is applied to one or more labelled training images in order to train a copy of said first machine learning classifier.

The invention discloses a rotating machine intelligent diagnosis method based on structured pruning and knowledge fusion distillation, and the method comprises the steps: training a teacher network through a training set, carrying out the structured pruning of a percentile threshold value on the teacher network based on the L2 norm calculation and normalization of a convolution filter, and generating a student network with a consistent structure. A KFD strategy including feature-level distillation and logit-level distillation is utilized to carry out deep supervision on a student network, and two types of distillation losses are weighted and fused, so that a student model still keeps relatively strong feature characterization capability and category discrimination capability under a high pruning rate. Asymmetric integer quantization is adopted for the trained student network, so that the reasoning overhead is reduced, and the embedded adaptability is improved. A general neural network operator IP core is arranged on an FPGA, efficient deployment of a quantitative student model is achieved, and low-power-consumption and low-delay real-time fault diagnosis is achieved. The method has the advantages of being high in precision, light in model weight, easy to deploy and the like, and is suitable for on-line monitoring of industrial field rotating machinery.

A rolling bearing fault feature extraction method based on variable scale multipoint kurtosis deconvolution comprises the following steps: a, sampling a rolling bearing vibration acceleration signal to obtain a rolling bearing fault vibration signal; b, constructing a toeplitz autocorrelation matrix; c, constructing a variable-scale multipoint kurtosis deconvolution filter; d, filtering the rolling bearing fault vibration signal by using an optimal deconvolution filter to obtain a fault impact signal; e, performing envelope demodulation processing on the fault impact signal, extracting an envelope of the fault impact signal, and obtaining an envelope spectrum through spectral analysis; and f, judging the fault type of the rolling bearing according to the envelope spectrum. According to the method, the optimal target vector of the deconvolution is searched by constructing the variable-scale multipoint kurtosis index, the optimal filter is constructed, the fault impact signal is extracted through the deconvolution, the fault impact signal is subjected to envelope analysis, the fault feature frequency of the rolling bearing is extracted, and the fault feature of the rolling bearing can be accurately extracted.

An example device for filtering video data includes a memory configured to store video data; and a processing system comprising one or more processors implemented in circuitry, the processing system being configured to: apply one or more neural network processing blocks to intermediate filtered video data, each of the neural network processing blocks including a first 1×1 convolutional filter, a parametric rectified linear unit (PReLU) filter, a second 1×1 convolutional filter, and a 3×3 convolutional filter; apply additional neural network processing blocks to output of the one or more neural network processing blocks to form filtered video data; and output the filtered video data.

In one embodiment, a method includes providing sequence information of a DNA sequence as an input to a neural network model, generating activations by convolutional filters of the neural network model based on the sequence information, wherein each convolutional filter is associated with a weight, identifying motifs from the DNA sequence based on the convolutional filters and their weights by the neural network model based on a regularization function configured to enable each convolutional filter to learn a distinct motif, generating a linear vector of attention scores for the motifs and identifying interactions between the motifs by attention layers of the neural network model based on the activations, determining motif instances and an associated syntax by the neural network model based on the linear vector of attention scores and the interactions, and generating predictions associated with genomic regulatory functions based on the motif instances and the associated syntax.

Systems and methods for image generation are provided. An aspect of the systems and methods includes obtaining a text prompt, generating a style vector based on the text prompt, generating an adaptive convolution filter based on the style vector, and generating an image corresponding to the text prompt based on the adaptive convolution filter.

The technology disclosed relates to a system that comprises a spatial convolution network and a temporal convolution network. The spatial convolution network is configured to process a window of per-cycle sequencing image sets and generate respective per-cycle spatial feature map sets. Trained coefficients of spatial convolution filters in spatial convolution filter banks of respective sequences of spatial convolution filter banks vary between sequences of spatial convolution layers in respective sequences of spatial convolution layers. The temporal convolution network is configured to process the per-cycle spatial feature map sets on a groupwise basis and generate respective per-group temporal feature map sets. Trained coefficients of temporal convolution filters in respective temporal convolution filter banks vary between temporal convolution filter banks in respective temporal convolution filter banks.

The invention discloses a radio frequency fingerprint identification method, device and equipment based on an enhanced complex convolutional neural network, and relates to the technical field of Internet of Things security, and the method comprises the steps: converting an I / Q signal transmitted by a to-be-identified radio frequency transmitter into a complex value radio frequency signal; and inputting the complex value radio frequency signal and the conjugate form thereof into the enhanced complex convolutional neural network to obtain an identification result of the radio frequency transmitter to be identified. A convolutional layer of a complex-valued convolutional block in the enhanced complex convolutional neural network is a wide linear complex-valued convolutional layer; and the wide linear complex value convolution layer performs convolution operation on the complex value radio frequency signal and the conjugate form thereof through two convolution filters with consistent parameters, and the processing results of the two convolution filters are added as the output of the wide linear complex value convolution layer. According to the method, the characteristics of the complex-valued radio-frequency signal are fully extracted through the wide linear complex-valued convolutional layer; and an SE attention mechanism module is introduced, so that the identification performance of the network model on the radio frequency fingerprint signal is effectively improved.

The invention relates to the technical field of construction site management, in particular to a highway construction site Internet of Things positioning and tracking system which comprises a fundamental frequency calculation module, a parameter generation module, a central frequency generation module, a central frequency calculation module and a central frequency calculation module. The fundamental frequency calculation module collects engine rotating speed and angular speed signals and calculates engine vibration fundamental frequency by combining ignition factors. The dynamic filtering module is used for configuring the weighting coefficient as a difference equation parameter, blocking bandwidth energy through convolution filtering and generating a purification angular velocity; the trajectory tracking module is used for calculating the purification angular velocity and a zero deviation value, calling a Runge-Kutta algorithm for integral generation of a course angle, and generating a position coordinate in combination with a driving speed. According to the method, the vibration fundamental frequency is inversely calculated by collecting the rotating speed of the engine, the blocking bandwidth is dynamically constructed according to the interference frequency, the frequency characteristic is mapped to be a difference equation coefficient, vibration noise is accurately stripped, course drift is eliminated, and it is ensured that the positioning coordinates matched with the actual track are output.

The invention provides an efficient image denoising method based on a hybrid filter technology, and belongs to the field of digital image processing. Traditional bilateral filtering and self-defined convolution filtering are fused to denoise an image; the method comprises the following steps: performing bilateral filtering on a noisy image for preprocessing, and performing self-defined convolution filtering on the preprocessed image to obtain a denoised image; the denoising precision can be improved while the real-time performance is ensured, and key pain points in the fields of industry, medical treatment, consumer electronics and the like are solved.

Disclosed are an FPGA-based method and device for surface defect detection of a surfaced workpiece. The method comprises: acquiring a magnetic induction intensity data matrix (S10); performing lock-in amplification on the magnetic induction intensity data matrix to obtain an original feature data matrix (S20); performing convolution filtering processing on the original feature data matrix to obtain a smoothed feature data matrix (S30); performing binarization processing on the smoothed feature data matrix to obtain a binarized data matrix (S40); on the basis of the binarized data matrix, determining whether magnetic field distortion is present on the surface of a surfaced workpiece (S50); and when magnetic field distortion is present on the surface of the surfaced workpiece, determining whether the magnetic field distortion is caused by a surface defect of the workpiece, and outputting a detection report on the basis of the determination result (S60). Defect detection is realized by means of hardware, fully utilizing the strong computational capability of the FPGA, and significantly improving the defect detection efficiency.

Provided are a method and an electronic device for performing convolutional neural network operations on a batch of input images using homomorphic encryption. The method includes: for each channel of the batch of the input images, grouping pixel values at the same location in the images included in the batch into a single vector, and encrypting the grouped vectors using a homomorphic encryption method to obtain a first ciphertext; for each channel and each coefficient of the ciphertext, grouping coefficients of the first ciphertext corresponding to all pixel locations to generate polynomials, and configuring the polynomials into a single matrix; multiplying the matrix by a polynomial matrix representing a convolution filter to obtain a result matrix, and performing a modulo operation on each element of the result matrix to generate an encrypted convolution operation result, the operation being performed on unencrypted matrices; and rearranging polynomial coefficients of the result matrix to generate a second ciphertext for the batch of the images after the convolution layer is applied, in which the second ciphertext is encrypted by the same homomorphic encryption method as the first ciphertext.

The invention belongs to the technical field of computers, and provides a Kolmogorov-Arnold convolution filter-based transfer learning method for human behavior recognition, and the method comprises the steps: firstly collecting and preprocessing sensor data of a wearable device, dividing the sensor data into a training set, a verification set and a test set, constructing a recognition model, and carrying out the recognition of the sensor data through a Kolmogorov-Arnold convolution filter. The method comprises the following steps: accessing a prompt module based on Kolmogorov-Arnold convolution to an input end of a pre-trained time sequence basic model with frozen backbone network parameters, accessing a classifier to an output end of the pre-trained time sequence basic model, then finely adjusting the prompt module and the classifier only by using training data, determining an optimal model through a verification set, and finally obtaining a time sequence basic model; and finally, the trained model is deployed on wearable equipment, real-time and high-precision human body behavior recognition is realized, and the conversion cost between different recognition tasks is greatly reduced.

The invention discloses an industrial defect detection method based on local attention enhancement and sparse query driving, and the method comprises the steps: improving an industrial image defect detection model of a DETR network, introducing a local attention mechanism for enhancement in a shallow feature extraction stage of a backbone network, fusing the enhanced shallow features with deep features, and carrying out the extraction of the deep features, and predicting a small target candidate frame based on the enhanced shallow features and initializing the small target candidate frame into a sparse query vector. In a Transform encoder and a Transform decoder, a dual-channel parallel module of a large kernel decomposition convolution combined convolution filter bank is adopted to replace a traditional self-attention mechanism. According to the invention, the problems of low precision and slow convergence rate of small-size defect detection are effectively solved, and high precision and automation of industrial defect detection are realized.

A computer model is trained with an architecture including additional training layers relative to the inference architecture. The architecture of a computer model to be used in inference includes a convolutional layer with a number of K×K convolutional filters. For training, the convolutional filters are expanded to a plurality of training layers including a layer with 1×1 and K×K filters. The expanded layers may include additional layers than the number of expanded filters in the layer of the inference model. The 1×1 expanded layer in training may learn weights for combining the K×K expanded layers, providing a weighted combination of the K×K filters for the respective channel of the layer of the inference layer.

The invention relates to the field of confocal microendoscopes, in particular to a fiber core bundle center positioning method and device of a confocal microendoscope and a terminal. The method comprises the steps of obtaining a plurality of frames of optical fiber end face images and performing fusion processing to generate a reference image; performing convolution filtering on the reference image based on the fiber core feature template to generate a fiber core response image; and positioning a plurality of candidate fiber core points of the fiber core response image, and clustering the plurality of candidate fiber core points to generate a fiber core beam estimation center. According to the method, noise in the image is filtered through fusion processing, filtering is performed through the fiber core feature template to highlight the fiber cores in the fused image, interference is filtered, and finally the estimation center of the fiber core bundle is obtained based on the distance clustering method, so that a background model is conveniently established for each fiber core and real-time tracking is performed on each fiber core in subsequent steps, and the accuracy of the estimation center is improved. According to the confocal microendoscope, the anti-interference capability is remarkably improved, the positioning precision is high, the response speed is high, and the imaging performance and reliability of the confocal microendoscope are improved.

The application provides a point spread function fast spatial convolution method, which comprises the following steps: step 1, calculating an imaging profile and a point spread function field; step 2, determining the spatial position coordinates of the center position of the point spread function by circulation; step 3, determining an interpolation method and calculating the interpolation coefficients of the local imaging profile which contributes to the point spread function; step 4, performing wave number domain fast spatial convolution; step 5, multiplying the profile by the corresponding interpolation coefficients after filtering; and step 6, accumulating the filtering results of the local imaging profile to obtain a final filtered profile. The point spread function fast spatial convolution method can greatly improve the calculation efficiency of the point spread function spatial convolution filtering process and reduce the calculation cost based on the linear accumulation characteristics of the interpolation method.

Systems and methods are disclosed. The method includes, for each of a plurality of spatial partial differential equations (sPDEs) within a wave equation, determining a linear system of equations using an approximate solution at a plurality of grid nodes. The linear system of equations includes an inverse matrix, a first vector, and a second vector. The method further includes, for each of the plurality of sPDEs, evaluating the inverse matrix by evaluating a first portion of the inverse matrix using a first deconvolution filter and evaluating a second portion of the inverse matrix using a second deconvolution filter. The method further still includes, for each of the plurality of sPDEs, evaluating the first vector using the evaluated inverse matrix and the second vector as well as determining the wavefield using the evaluated first vector for each of the plurality of sPDEs, a velocity model, and the wave equation.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing an input image using a super-resolution neural network to generate an up-sampled image that is a higher resolution version of the input image. In one aspect, a method comprises: processing the input image using an encoder subnetwork of the super-resolution neural network to generate a feature map; generating an updated feature map, comprising, for each spatial position in the updated feature map: applying a convolutional filter to the feature map to generate a plurality of features corresponding to the spatial position in the updated feature map, wherein the convolutional filter is parametrized by a set of convolutional filter parameters that are generated by processing data representing the spatial position using a hyper neural network; and processing the updated feature map using a projection subnetwork of the super-resolution neural network to generate the up-sampled image.

The invention discloses a multi-intelligent reflecting surface (IRS) channel prediction method based on transfer learning, and belongs to the technical field of wireless communication. The method comprises the following steps: constructing a multi-IRS system model, representing channel characteristics, constructing a data set, designing a physical prior convolutional neural network (Physics-aware CNN), pre-training a source domain, adapting a target domain, predicting the channel and evaluating the performance. By introducing a frequency domain convolution filter and a phase alignment layer and combining a migration strategy of layer-by-layer unfreezing, the data demand of a target domain is remarkably reduced, and the prediction precision is improved. According to the method, the approximate optimal performance can be achieved when only 30% of target domain data is used, the pilot frequency overhead is effectively reduced, low errors are kept in a large-scale high-dimensional channel scene, and good robustness and engineering application value are achieved.

To provide a portable and embedded type system and an artificial intelligence based method which arrange a deep Convolution Neural Network (CNN).SOLUTION: The spatial convolution network is configured to: process a window of per-cycle sequencing image sets; and generate respective per-cycle spatial feature map sets. Trained coefficients of spatial convolution filters in spatial convolution filter banks of respective sequences of spatial convolution filter banks vary between sequences of spatial convolution layers in respective sequences of spatial convolution layers. The per-cycle spatial feature map sets are processed on a groupwise basis, and respective per-group temporal feature map sets are generated. Trained coefficients of temporal convolution filters in respective temporal convolution filter banks vary between temporal convolution filter banks in respective temporal convolution filter banks.SELECTED DRAWING: Figure 25A

The invention discloses a vibration suppression method, device and equipment and a readable storage medium, and relates to the technical field of vibration suppression, and the vibration suppression method comprises the steps: obtaining an operation feedback signal in a process that industrial equipment operates according to a position instruction; extracting a motor vibration signal in the operation feedback signal; determining a vibration suppression parameter according to the motor vibration signal; and according to the vibration suppression parameters, performing convolution filtering processing on the initial position instruction to obtain a vibration suppression position instruction, and assigning the vibration suppression position instruction to the position instruction. The problem that due to the low-frequency vibration phenomenon caused by the characteristics of a flexible transmission part, the stability performance of industrial equipment during operation is reduced is solved.

The application discloses a curve structure image enhancement method and device based on filtering and geometric constraint interaction, and relates to the technical field of image enhancement. The method comprises the following steps: a to-be-processed degraded image is subjected to multi-scale feature coding through a backbone network and a feature pyramid network to obtain feature maps of multiple levels. Linear / curve structure regions to be repaired are determined on the feature maps of the multiple levels based on structure guiding information, and corresponding region-of-interest features are intercepted. Spatial filtering is performed on the region-of-interest features by using a large-scale convolution filter to realize global context perception processing to repair texture breakage caused by occlusion, and repaired features are obtained. A pixel-level self-attention mechanism is used to perform geometric feature interaction and smoothing filtering on a curve sampling point sequence represented by the repaired features to suppress high-frequency geometric noise, and smoothed features are obtained. Image texture reconstruction is performed on the smoothed features to generate an enhanced image containing complete curve texture.

This application provides a method, apparatus, device, and storage medium for cell nucleus segmentation based on contour characteristics. The method includes: acquiring an image to be segmented; preprocessing the image to be segmented to obtain a preprocessed image; using a convolutional filter on the preprocessed image to determine a denoised image; determining effective cell nucleus contours based on the denoised image; and expanding the effective cell nucleus contours based on adjacent pixels to obtain segmented cell nuclei. This scheme can achieve high accuracy on real cervical cell images (i.e., the BSMMU dataset) while maintaining a fairly high recall rate.

The technology disclosed relates to constructing a convolutional neural network-based classifier for variant classification. In particular, it relates to training a convolutional neural network-based classifier on training data using a backpropagation-based gradient update technique that progressively match outputs of the convolutional network network-based classifier with corresponding ground truth labels. The convolutional neural network-based classifier comprises groups of residual blocks, each group of residual blocks is parameterized by a number of convolution filters in the residual blocks, a convolution window size of the residual blocks, and an atrous convolution rate of the residual blocks, the size of convolution window varies between groups of residual blocks, the atrous convolution rate varies between groups of residual blocks. The training data includes benign training examples and pathogenic training examples of translated sequence pairs generated from benign variants and pathogenic variants.

The present invention discloses an image classification method based on the fusion of redundancy and diversity features, including: replacing some filters in the convolutional neural network model with extended fusion convolutional filters to form a lightweight convolutional network model; training and validating the lightweight convolutional network model using a data set to obtain a trained lightweight convolutional network model; inputting the image to be processed into the trained lightweight convolutional network model to obtain a classification result. The lightweight convolutional network model can not only obtain inherent features with a certain degree of redundancy, but also generate diversity features containing necessary detailed information, improving the accuracy and feature diversity of the model while effectively reducing the number of model parameters and floating-point operations of calculations, thereby enhancing the classification effect; in addition, the extended fusion convolutional filter can also be used as a plug-and-play component to upgrade the existing convolutional neural network, and the hyperparameters are globally and uniformly set to facilitate the rapid construction of the model, making the model have better performance.

The embodiment of the application discloses a shadow generation method and related device, the method comprises the following steps: obtaining the depth data, normal data and illumination data of each point in a virtual scene as the shadow reference data of each point; inputting the shadow reference data of each point into a first convolution filter to obtain the basic shadow data of each point output by the first convolution filter; determining the half-shadow intensity of each point according to the distance from each point to a virtual light source in the virtual scene and the propagation distance of a virtual light emitted by the virtual light source; weighting the basic shadow data of each point by using the half-shadow intensity of each point to obtain the weighted shadow data of each point; and the weighted shadow data is used for shadow rendering for the points. The method can improve the setting efficiency of the shadow effect, and reduces the required artificial cost and time cost.