36 results about "Feature compensation" patented technology

The invention discloses a feature compensation method based on rapid noise estimation in a speech recognition system. The method is characterized in that noise parameter estimation in the feature compensation is separated from pure speech estimation, and noise estimation and pure speech estimation are achieved through different Gaussian mixture models (GMMs). A GMM containing less Gaussian units is used for extracting noise parameters from a noisy tested speech; another GMM containing more Gaussian units is used for being combined with an estimated single Gaussian noise model to obtain a noisy GMM matched with the current test environment; and finally the noisy GMM is used for calculating the posterior probability of the noisy tested speech and the pure speech feature vector is estimated from the noisy tested speech through the minimum mean square error method. According to the method, estimation accuracy of the pure speech can be guaranteed while the calculated amount is reduced.

The invention relates to a sound-groove identification method based on emotion compensation. The emotion compensation includes three portions of emotion detection, character compensation and emotion expansion, comprising of calculating voice emotion factors to be according to the emotion detecting technique, compensating the voice change caused by emotion change respectively from the two layers of character and mode and finally improving robustness of the sound-groove identification technique to the emotion change. The invention has the advantages that the invention breaks through the inconsideration of sound-groove emotion change of the existing sound-groove identification technique, deals with the voice change caused by emotion change from the two layers of character and mode and strengthens resisting power to the voice emotion drift. The character layer standardizes the voice feature within the modeling ability of the training model by means of emotion degradation, normalization and barrier to reach the purpose of inhibiting the influence of the user emotion on the identification property. The mode layer obtains large scale emotion voices by employing the reverse way of synthesizing emotion voice by emotion changing rule, thereby greatly expanding the modeling information of the sound-groove model and resoling the difficulty of obtaining emotion data.

The invention discloses an isolation word identification method based on a double-layer GMM structure and VTS feature compensation. The method comprises a training stage and an identifying stage. In the training stage, by voice feature extracting under a pure environment, two GMM training models and an HMM training models are obtained. Each GMM model comprises a GMM1 model containing a small number of Gauss mixing units and a GMM2 model containing a large number of Gauss mixing units. During a noise estimation process at a vector Taylor series (VTS) feature compensation stage, the GMM1 model is used for obtaining the mean value and the variance of noise, a GMM2 model is used for obtaining a pure feature parameters by mapping, and matching with the HMM module is carried out to obtain the final identification results. Compared with an isolation word identification algorithm based on a single GMM model and VTS feature compensation, under the situation that the error recognition rate is not changed basically, noise mean value and variance estimating time is shortened by 90%, feature compensation overall time is shortened by 30%-50%, and calculated quantity of the isolation word identification algorithm based on the VTS feature compensation is effectively lowered.

The invention provides a behavior identification method for long-time fast and slow network fusion based on attitude articulation points. The method comprises the following steps: automatically capturing modes of articulation point structure features and track features in a space and time sequence by utilizing a graph convolutional network; generating an overall space-time feature of each video clip through a feature splicing network model, connecting the overall space-time features in series according to a clip sequence to form the overall space-time features of the video, and fusing the RGBfeatures and the attitude joint point features extracted from the input video at a high level of the convolutional network; and outputting the classification result of the video behavior through a support vector machine classifier in a weighted fusion manner. By extracting the features of the attitude joint points, the data size is greatly reduced, and redundant information is removed. Meanwhile,the space-time features in the long-time-history multi-frame image sequence are extracted to carry out feature compensation, so that the recognition rate and robustness of the video complex behavior are improved.

The invention relates to a speech emotion recognition method based on a multistage residual convolutional neural network, and belongs to the technical field of speech signal analysis, image processingand the like. The method comprises the following steps: 1) a training process: collecting and preprocessing sound signals with all emotions to generate a spectrogram; constructing a multi-stage residual convolutional neural network, and inputting the spectrogram into the multi-stage residual convolutional neural network for training; 2) a test process: acquiring and preprocessing a to-be-identified sound signal, and generating a to-be-identified spectrogram; and then inputting the to-be-identified spectrogram into the trained multistage residual convolutional neural network to obtain a recognition result. According to the method, the CNN is subjected to feature compensation by crossing multi-stage residual blocks, so that the problem of feature loss of the CNN along with deepening of a convolution layer is solved, and the recognition rate is increased.

The embodiment of the invention discloses an image processing method and device, an electronic device and a storage medium, and the method comprises the steps: obtaining a first image frame sequence which comprises a to-be-processed image frame and at least one adjacent image frame adjacent to the to-be-processed image frame; performing image alignment operation on the to-be-processed image frameand each adjacent image frame to obtain alignment feature data of each adjacent image frame; and performing feature compensation operation on the alignment feature data of each adjacent image frame based on the to-be-processed image to obtain compensated alignment feature data of each adjacent image frame, and determining a processed image frame corresponding to the to-be-processed image frame based on the compensated alignment feature data. In the embodiment of the invention, after the preliminary image alignment operation is carried out on the multiple image frames, the feature compensationoperation is carried out on the image frames after the preliminary image alignment operation is carried out, so that the image alignment and fusion quality is effectively improved.

The invention relates to a display panel and a compensation data transmission method, and the method comprises the steps: generating feature compensation data; transmitting the feature compensation data to a first storage unit; transmitting the feature compensation data stored in the first storage unit to a second storage unit. According to the embodiment of the invention, the compensation data transmission method does not need a Flash memory. A screen module group does not need to be provided with the Flash memory, thereby reducing the physical space occupied by the Flash memory in the screenmodule group. In the embodiment of the invention, the method can generate the feature compensation data through the Mura compensation data, and the feature compensation data is directly stored in a data drive circuit. The method reduces the error codes in the transmission process of the feature compensation data. The method does not need an additional Flash memory on the screen module group, thereby reducing the data transmission link, improving the data transmission efficiency, and reducing the design cost of the display panel.

The invention discloses a feature compensation video target tracking method based on a posterior pixel color histogram, a direction gradient histogram and a convolutional neural network, which uses simple features in a simple scene to ensure real-time performance, and uses complex features in a complex scene to ensure accuracy. Through combination of two features of a posterior pixel histogram anda direction gradient histogram, an obtained response feature map can well adapt to a simple condition of a video scene. Training a classifier to judge which situation the response obtained by the former is fused is untrusted; and finally, according to the judgment result of the classifier, selecting whether to add a convolutional neural network tracker which is slow in speed and more robust in performance to correct the target deviated from the tracking, or re-finding the target lost from the tracking. According to the method, the precision of judging the size and the position of the target in the video is improved, and the method can well adapt to a long-time target tracking task to achieve a scene of practical application.

A system and method of adjusting mask feature dimensions to compensate for anomalies in energy output of a laser beam used in ablating a workpiece. This method involves measuring the effects of variations in energy output of the laser beam at several points in the laser beam profile. In the case where the workpiece is a nozzle plate, the features adjusted in the mask are the diameters of the mask holes. Differences in the effects of the energy output between the points in the laser beam are then compensated for by adjusting the mask features that correspond to the points in the laser beam.

The embodiment of the invention discloses an audio processing method, related equipment, a storage medium and a program product, and the method comprises the steps: obtaining to-be-processed target audio data and the frequency spectrum information of the target audio data,wherein the audio missing clip exists in the target audio data, the frequency spectrum information comprises frequency spectrum features of context audio clips of the audio missing clip; performing feature compensation processing on the frequency spectrum information of the target audio data according to the frequency spectrum features of the context audio clip to obtain compensation frequency spectrum information corresponding to the target audio data; performing audio prediction by adopting the compensation frequency spectrum information to obtain predicted audio data; and compensating the audio missing clip in the target audio data according to the predicted audio data to obtain the compensated audio data of the target audio data, so that the accuracy of carrying out prediction compensation on the packet loss audio can be improved.

The invention relates to a multichannel speech recognition acoustic modeling method and device based on spatial feature compensation. The model is based on a traditional mixing acoustic modeling frame, namely, a neural network acoustic model, the state posterior probability of a Hidden Markov Model is predicted, and the method comprises the steps that the acoustic feature of speech signals recorded by each single channel in a microphone array and space information features in the microphone array are extracted; the acoustic feature and the space information features are input into the neural network acoustic model to be trained; the predicted acoustic state posterior probability is output by the neural network acoustic model, an acoustic model optimization criterion is used for carrying out iterative updating on neural network parameters, and a multichannel speech recognition acoustic model based on spatial feature compensation is generated. According to the method, a second-best solution caused by separate optimization of the front end and the rear end in a traditional method is avoided; space information provided by the microphone array is effectively utilized by the neural network acoustic model, and the acoustic modeling capacity on multichannel speech signals is improved.

The invention discloses a robust speech recognition method based on multi-band characteristic compensation, comprising: decomposing MFCC (Mel frequency cepstrum coefficient) extracted in noise environment directly in a cepstrum domain to obtain sub-MFCCs of four bands; then, using a pre-trained Gaussian hybrid model on each band to perform characteristic compensation on the sub-MFCC of the present band to obtain pure sub-characteristic parameters; finally, subjecting the estimated sub-MFCCs of the four bands to characteristic combination to obtain complete MFCC of pure speech. The method can provide improved recognition performance for a speech recognition system in a limited noise environment, and improved noise robustness for the system.

The invention discloses a space target autonomous identification method based on transfer learning, electronic equipment and a storage medium. The method comprises the steps of obtaining a plurality of frames of image data of an unknown space target in real time; performing feature compensation on the image data of the unknown space target at the current moment according to the image data of the unknown space target at a plurality of frames of preorder moments to obtain an unknown space target feature map at the current moment after feature compensation; and recognizing the unknown space target feature map subjected to feature compensation at the current moment by adopting a pre-trained neural network to obtain a recognition result. According to the method, the problem that the number andtypes of samples are not rich enough due to loss of priori information of a non-cooperative space target is solved.

The invention discloses a dialectal context multi-language translation method, belongs to the computer language translation technology field and solves a problem that an error translation rate increases as a language translation tool in the prior art can not accurately translate dialects. The method comprises steps that step 1, a characteristic code database formed by characteristic codes of various types of dialect voices under a native language system is established; step 2, dialect characteristic compensation codes are generated according to difference between the characteristic codes and standard base voice characteristic codes, and a characteristic compensation code database is formed; step 3, the characteristic compensation codes and corresponding source voices are compounded to generate standard base voices under the native language system; and step 4, the standard base voices are converted into voices or texts of a target language. The method is applied to translation of dialects and customized voices.

The invention discloses a compensation algorithm for three-dimensional space measurement result data, and the algorithm comprises the steps: obtaining point cloud data: scanning an exposed surface ofan object through employing laser radar equipment, obtaining initial data, and converting the initial data through employing a trigonometric function to obtain rectangular coordinate system data; compensating the bottom point cloud of the object, processing the coordinates by using algorithm software based on the obtained original point cloud coordinates, and keeping the coordinate values of the original point cloud unchanged to obtain bottom point cloud data; compensating point clouds around the object, finding out a difference value between a minimum value coordinate and a maximum value coordinate in the point clouds through a comparison algorithm according to bottom surface point cloud data, and screening out projection area edge points of the minimum value and the maximum value of eachsection through the comparison algorithm. The invention has the technical effects that the missing surface point cloud is calculated through feature compensation of the surface of the target object,manual compensation errors are reduced, the point cloud of the missing surface is compensated according to the feature density of the target object, and approximately closed complete point cloud dataof the target object is obtained.

The invention relates to an artificial limb socket design quantization compensation method based on a feature vector method. The method comprises the steps that residual limb tissue influence factorsare designed by an artificial limb socket; according to individual information weight design in artificial limb specialist weight design, the specialist employment period, the technical grade and thelike for artificial limb socket design are simplified into quantizable indexes, a judgment matrix is calculated, and a feature vector corresponding to the largest eigenvalue is obtained as an estimated individual information weight vector; the comprehensive evaluation weight vector of the artificial limb specialist is obtained by multiplying the individual information weight vector by a specialistindividual information parameter index quantification matrix; the artificial limb socket soft tissue feature compensation index weight is obtained by summing a socket quantization compensation simplification table of each specialist and the weight product, and the socket quantization compensation simplified table is obtained by simplifying the socket design influence factor quantization compensation simplification table according to simplification standards; the artificial limb socket tissue feature compensation quantization model is obtained according to a compensation value corresponding tothe compensation index in the socket design influence factor compensation simplification standard.

A feature compensation apparatus includes a feature extractor configured to extract corrupt speech features from a corrupt speech signal with additive noise that consists of two or more frames; a noise estimator configured to estimate noise features based on the extracted corrupt speech features and compensated speech features; a probability calculator configured to calculate a correlation between adjacent frames of the corrupt speech signal; and a speech feature compensator configured to generate compensated speech features by eliminating noise features of the extracted corrupt speech features while taking into consideration the correlation between adjacent frames of the corrupt speech signal and the estimated noise features, and to transmit the generated compensated speech features to the noise estimator.

The invention discloses a method for noise suppression by using multiple digital pictures, which is used for denoising by virtue of multiple digital pictures, wherein the digital pictures are firstly subject to a feature weighting processing program and the picture feature compensation of target pixels, and then noise suppression is carried out on the target pixels by carrying out interactive reference on multiple continuous or similar digital pictures. The method for noise suppression by using multiple digital pictures, which is provided by the invention, comprises the following steps: acquiring multiple digital pictures; carrying out a first pixel compensation program on each digital picture; and carrying out a second pixel compensation program on a comparison picture in the digital pictures. In the noise suppression method disclosed by the invention, pixels of which the picture features are similar and the degree of similarity is high are used to repair, thus not destroying the digital pictures, so as to obtain better picture quality.

The invention belongs to the technical field of remote sensing images, and discloses a remote sensing scene classification method and device, terminal equipment and a storage medium. The method comprises the steps of obtaining a remote sensing scene image set, and inputting the remote sensing scene image set into a preset convolutional neural network model for feature extraction to obtain a top semantic feature set and a shallow appearance feature set; performing feature aggregation on the top semantic feature set through dense connection to obtain a first convolution feature; performing feature aggregation on the shallow appearance feature set to obtain a second convolution feature; performing feature compensation on the first convolution feature and the second convolution feature throughbidirectional gating connection to obtain a target convolution feature; and classifying the remote sensing scene images in the remote sensing scene image set according to the target convolution features. Characteristic aggregation is utilized, and shallow appearance characteristics and top semantic characteristics are complemented, so that shallow convolution characteristic information loss in aclassification characteristic aggregation stage is prevented.

The invention discloses an underwater target classification method considering robustness of a deep learning model, and aims to solve the problem that the existing deep learning model is low in underwater target classification accuracy. The method comprises the following steps: predicting a collected underwater target data training set by using a trained original model to obtain a set of all correctly classified samples and a set of all wrongly classified samples; inputting the set of all the classified error samples into the trained original model, and performing clustering and feature compensation on the features of the classified error samples to obtain feature compensation of the classified error samples; inputting feature compensation into the trained original model to obtain an original model after feature compensation; inputting the underwater target data training set into the original model after feature compensation, and outputting samples with wrong classification; building an adversarial training model to obtain a trained adversarial training model; performing weighted combination on the adversarial training model and the original model after feature compensation to generate a deep learning model; the method belongs to the underwater target classification field.

The present application discloses an accuracy compensation method for feature mapping scaling, which includes the following steps: obtaining mapping data of original features and target features, wherein the mapping data at least includes bit width, number of channels, number of feature horizontal pixels and feature vertical pixels Number; according to the original feature map and the number of feature horizontal pixels and the number of feature vertical pixels of the target feature, calculate the target feature in the horizontal direction and the number of feature vertical pixels based on the bit width on each channel indicated by the channel number. Interpolation coordinates in the vertical direction; according to the respective interpolation coordinates of the original feature map and the target feature, calculate the interpolation of the original feature at each interpolation coordinate of the target feature based on the bit width on each channel indicated by the channel number Weight; determine the target feature map according to the pixel value of the original feature at each position and the interpolation weight at each interpolation coordinate. The application also discloses the corresponding computer system and storage medium.

According to the one-stage safety helmet detection method based on the S3FD network, on the basis of a one-stage detector S3FD, the S3FD is changed into a network comprising a filtering structure, a connection structure and an accurate judgment structure. The network can filter redundant samples, has promotion significance for small target detection, and can improve the detection capability of the safety helmet to a great extent. In addition, a feature compensation module is provided, and the module can reduce feature loss in the convolution operation process. Therefore, the problem of few effective features in small target detection is solved, and the safety helmet detection accuracy is improved to a certain extent. Through the improvement, the training speed of the model is accelerated, and the training efficiency and the detection precision are both improved.

The invention provides an indoor positioning method, related equipment and a computer readable storage medium, which are used for improving the positioning precision of terminal equipment on the basis of increasing the quantity of channel feature information. The method comprises the following steps: when a first spatial position of terminal equipment is determined to be known, updating a channel feature library based on acquired multiple pieces of first channel feature information; training is carried out based on the updated channel feature library, and the number of channel feature information included in the updated channel feature library is increased; acquiring first channel feature compensation information corresponding to the channel feature library after the quantity of the channel feature information is increased based on the channel feature library after the quantity of the channel feature information is increased; and the positioning server or the terminal device positions the terminal device based on the channel feature library after the number of the channel feature information is increased and the first channel feature compensation information.

The invention discloses a multi-spectrogram fusion method and device for realizing human body posture estimation based on millimeter wave radar, and the method comprises the steps: obtaining a distance-Doppler spectrogram and a distance-angle spectrogram, carrying out the attention matrix calculation of the two obtained spectrograms through a Transform structure, and obtaining the fusion features of the two spectrograms; repairing the fusion features by using the distance-Doppler spectrogram features to obtain complete fusion features; obtaining time sequence features of continuous frames by using a bidirectional long short-term memory network, and performing feature compensation on the time sequence to obtain rough compensation features; for the rough compensation features of each frame, calculating relevance between human joints in the same frame and performing spatial feature compensation according to the relevance to obtain fine compensation features of each frame; and mapping the fine compensation feature of each frame into a human skeleton to complete human posture estimation of the millimeter wave radar. According to the invention, multi-spectrogram fusion is effectively realized, and the accuracy of human body posture estimation realized by the millimeter wave radar is greatly improved.